Technology Evidence File
May21: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
The BBC and other media broke a story about China’s large scale testing of emotion detecting facial recognition/AI systems in Xinjiang, where it has been accused of genocide against the Muslim Uighur population | According to the BBC story (“AI Emotion-Detection Software Tested On Uyghurs”, by Jane Wakefield), an engineer familiar with the system claimed “the AI system is trained to detect and analyse even minute changes in facial expressions and skin pores…the software creates a pie chart, with the red segment representing a negative or anxious state of mind. The engineer claimed the software was intended for pre-judgment without any credible evidence". Apparently, in their pursuit of ever more effective surveillance and social control technologies, Chinese authorities are not concerned with findings reported in a recent Financial Times editorial, “Computers Are Not The Best Judge Of Our Emotions”. The FT notes that, “emotional AI is being used in sectors ranging from advertising to gaming to insurance, as well as law enforcement and security. It holds out the prospect of using facial clues to figure out what to sell people and how they respond to adverts; to check whether drivers or schoolchildren — or those working from home — are paying attention; or to spot people who are acting suspiciously… “Trials have suggested AI systems are far from perfect at discerning whether humans are telling the truth… Critics charge that the underlying assumptions of emotion recognition — that humans experience a universal set of emotions and manifest them in similar ways — are deeply flawed, and ignore differences between cultures and even individuals… “If it can be made to function in a reliable and trusted manner, emotion recognition offers at its best a way to humanize technology and to help businesses to understand customers far more deeply. At worst, it could be an invasive tool of surveillance capitalism. The right balance must be found.” |
“Deep Tech: The Great Wave of Innovation”, by BCG | Get ready for the great wave—the next big surge of innovation powered by emerging technologies and the approach of deep tech entrepreneurs. Its economic, business, and social impact will be felt everywhere because deep tech ventures aim to solve many of our most complex problems… “Deep tech describes an approach enabled by problem orientation and the convergence of approaches and technologies, powered by the design-build-test-learn (DBTL) cycle that de-risks and accelerates both product development and time to commercialization… The great wave encompasses artificial intelligence (AI), synthetic biology, nanotechnologies, and quantum computing, among other advanced technologies. But even more significant are the convergences of technologies and of approaches that will accelerate and redefine innovation for decades to come.” |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
Apr21: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Combating Ransomware”, Final Report of the Ransomware Task Force Note that this report was published before the recent ransomware attack that shut down the Colonial Pipeline from Gulf Coast to the Northeast US | SURPRISE “Ransomware attacks present an urgent national security risk around the world. This evolving form of cybercrime, through which criminals remotely compromise computer systems and demand a ransom in return for restoring and/or not exposing data, is economically destructive and leads to dangerous real-world consequences that far exceed the costs of the ransom payments alone… “Despite the gravity of their crimes, the majority of ransomware criminals operate with near-impunity, based out of jurisdictions that are unable or unwilling to bring them to justice. “This problem is exacerbated by financial systems that enable attackers to receive funds without being traced. “Additionally, the barriers to entry into this lucrative criminal enterprise have become shockingly low. “The “ransomware as a service” (RaaS) model, allows criminals without technical sophistication to conduct ransomware attacks. At the same time, technically knowledgeable criminals are conducting increasingly sophisticated attacks. “Significant effort has been made to understand and address the ransomware threat, yet attackers continue to succeed on a broad and troubling scale. To shift these dynamics, the international community needs a comprehensive approach that influences the behavior of actors on all sides of the ecosystem, including deterring and disrupting attackers, shoring up preparation and response of potential victims, and engaging regulators, law enforcement, and national security experts. “We also need international cooperation and adoption of processes, standards, and expectations… “While we have identified some recommendations as priorities, we strongly recommend viewing the entire set of recommendations together, as they are designed to complement, and build on each other. “The strategic framework is organized around four primary goals: to deter ransomware attacks through a nationally and internationally coordinated, comprehensive strategy; to disrupt the business model and reduce criminal profits; to help organizations prepare for ransomware attacks; and to respond to ransomware attacks more effectively.” |
“Sun Tzu Versus AI: Why Artificial Intelligence Can Fail in Great Power Conflict”, by Captain Sam Tangredi, US Navy (Retired) | “Recent Department of Defense (DoD) officials—following the thinking of political and corporate leaders—appear uniformly to perceive (or at least state rhetorically) that AI is making fundamental and historic changes to warfare. Even if they do not know all it can and cannot do, they “want more of it.” “While this desire to expand military applications of AI as a means of managing information is laudable, the underlying belief that AI is a game-changer is dangerous, because it blinds DoD to the reality that today’s battle between information and deception in war is not fundamentally, naturally, or characteristically different from what it was in the past. It may be faster; it may be conducted in the binary computer language of 1s and 0s; it may involve an exponentially increasing amount of raw data; but what remains most critical to victory is not the means by which information is processed, but the validity of the information… “That needs to be emphasized: More data will be false—in a similar fashion to the on-the-spot opinions, tweets, and conspiracy theories that increasingly clog social media. This is a significant problem for both DoD and AI development. “Much of the AI used in the corporate world—particularly concerning internet generated data—has been created with no concern about deliberate deception. However, if one assumes that a potential customer is likely to deliberately deceive a supplier as to the products or services he or she might buy, then the whole model of AI-assisted marketing crumbles. “If one assumes that companies within a supply chain might deceive the assembler as to their part specifications, the supply chain cannot function. If the data is false, AI is no longer a commercial asset for marketing or production—it becomes a liability. “For commercial AI to function, it must assume that it is not being deceived… “Whoever becomes the leader in this sphere will become ruler of the world.” The reality is that it is not the one who has the “best AI” who will dominate politico-military decision-making, but the one—all other elements of power being relatively equal—who has the most accurate, meaningful, and deception-free data.” |
As the EU proposed new rules on the use of Artificial Intelligence, a case in the UK showed why they are needed. | As the Financial Times’ John Thornhill decribes, in the UK the Criminal Court of Appeal “quashed the conviction of 39 sub-postmasters for theft, fraud, and false accounting” based on evidence provided by a flawed computer system (“Horizon”) that tracked their accounts. “The judgment clears the way for many of the other 700 sub-postmasters prosecuted using evidence from the Horizon system to challenge their convictions” … Thornhill notes that, “from a legal perspective, the affair highlights the dangers of humans blindly accepting the output of automated systems as reliable evidence, the computer never lies mentality” (“Post Office Scandal Exposes the Risk of Automated Injustice”). The proposed EU AI regulations are groundbreaking. In “The EU Path Towards Regulation of Artificial Intelligence”, Marcia and Desouza from the Brookings Institution, explain that the EU proposal “differentiates the uses of AI according to whether they create an unacceptable risk, a high risk, or a low risk. The risk is unacceptable if it poses a clear threat to people’s security and fundamental rights and is prohibited for this reason. “The European Commission has identified examples of unacceptable risk as uses of AI that manipulate human behavior and systems that allow social-credit scoring. For example, this European legal framework would prohibit an AI system similar to China’s social credit scoring. “The European Commission defined high-risk as a system intended to be used as a security component, which is subject to a compliance check by a third party. “The concept of high-risk is better specified by the Annex III of the European Commission’s proposal, which considers eight areas. Among these areas are considered high-risk AI systems related to critical infrastructure (such as road traffic and water supply), educational training (e.g., the use of AI systems to score tests and exams), safety components of products (e.g., robot-assisted surgery), and employees’ selection (e.g., resume-sorting software). “AI systems that fall into the high-risk category are subject to strict requirements, which they must comply with before being placed on the market. Among these are the adoption of an adequate risk assessment, the traceability of the results, adequate information on the AI system must be provided to the user, and a guarantee of a high level of security. Furthermore, adequate human control must be present.” |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
Mar21: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“People Systematically Overlook Subtractive Changes”, by Adams et al | SUPRISE In his book, “The Collapse of Complex Societies”, Joseph Tainter provided evidence for his theory that as societies solve problems, they become more complex over time. However as complexity increases, it produces diminishing positive returns, and later increasing negative returns. Eventually, this process leads to collapse. In addition, we noted that growing complexity also leads to increasing uncertainty, which also has a range of negative effects, such as increased conformity, reduced investment and other social phenomena. In this paper, the authors show that the tendency towards increasing complexity has deep psychological roots, finding that, “people systematically default to searching for additive transformations [that increase complexity], and consequently overlook subtractive transformations [that decrease it].” |
“Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks”, by Northcutt et al | SUPRRISE “Large labeled data sets have been critical to the success of supervised machine learning across the board in domains such as image classification, sentiment analysis, and audio classification. Yet, the processes used to construct datasets often involve some degree of automatic labeling or crowdsourcing, techniques which are inherently error-prone… “Researchers rely on benchmark test datasets to evaluate and measure progress in the state-of-the-art and to validate theoretical findings. If label errors occurred profusely, they could potentially undermine the framework by which we measure progress in machine learning … “We identify label errors in the test sets of 10 of the most commonly-used computer vision, natural language, and audio datasets, and subsequently study the potential for these label errors to affect benchmark results. Errors in test sets are numerous and widespread: we estimate an average of 3.4% errors across the 10 datasets” … As a result, “higher-capacity [machine learning] models undesirably reflect the distribution of systematic label errors in their predictions to a far greater degree than models with fewer parameters, and this effect increases with the prevalence of mislabeled test data.” |
“Accelerating Science with Human Versus Alien Artificial Intelligences” by Sourati and Evans | SURPRISE “Research across applied science and engineering, from materials discovery to drug and vaccine development, is hampered by enormous design spaces that overwhelm researchers’ ability to evaluate the full range of potentially valuable candidate designs by simulation and experiment. “To face this challenge, researchers have initialized data-driven AI models with published scientific results to create powerful prediction engines. “These models are being used to enable discovery of novel materials with desirable properties and targeted construction of new therapies. But such efforts typically ignore the distribution of scientists and inventors—human prediction engines—who continuously alter the landscape of discovery and invention. “As a result, AI algorithms unwittingly compete with human experts, failing to complement them and augment collective advance. “As we demonstrate, incorporating knowledge of human experts and expertise can improve predictions of future discoveries by more than 100% above AI methods that ignore them. “Nevertheless, with tens of millions of active scientists and engineers around the world, is the production of artificial intelligences that mimic human capacity our most strategic or ethical investment? “By not mimicking, but rather avoiding human inferences we can design “alien” AIs that radically augment rather than replace human capacity. Identifying the bias of collective human discovery, we demonstrate how human-avoiding or alien algorithms broaden the scope of things discovered by identifying hypotheses unlikely for scientists and inventors to imagine or pursue with undiminished signs of scientific and technological promise… not only accelerating but punctuating scientific advance.” |
“Morningstar Unleashes Robots to Write Fund Research”, by Michael Mackenzie in the Financial Times | “Morningstar has found a way to increase its written research without further taxing its army of human analysts. “The machine-generated reports that began rolling out this week set out the rationale behind Morningstar’s so-called analyst rating on a fund… Morningstar said this week that the robot ratings have performed as well as the recommendations generated by human analysts, based on three years of data.” |
| |
| |
| |
| |
| |
| |
| |
| |
| |
Feb21: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Super Mario meets AI: The Effects of Automation on Team Performance and Coordination in a Videogame Experiment”, by Dell’Acqua et al | SURPRISE “Recent advances in artificial intelligence (AI) have piqued interest in how these technological advances will transform jobs and labor markets. While prior work has focused on understanding the tasks where AI outperforms humans, we ask how the introduction of automated agents affects teams, their routines, and organizations… “We demonstrate experimentally that even in a task where automated agents outperform humans, the introduction of an automated agent decreases team performance. These effects are especially large in the short-term and in low- and medium-skilled teams. “We furthermore document that automation can generate adverse spillover effects into teams that do not receive an automated agent but must coordinate with it. “Our results indicate that these effects are driven by an increase in coordination failures, and we provide suggestive evidence that automation reduces team trust and individual effort provision. “Overall, our team-based approach highlights that human-machine interaction is key to expanding our understanding of how AI will transform teams, organizations, and work more broadly.” |
“Deepfake is the Future of Content Creation”, by Bernd Deubusmann of the BBC | This is another indication that uncertainty about information integrity will very likely exponentially increase in the coming years. “When most people think of deepfakes, they imagine fake videos of celebrities…Despite the negative connotations surrounding the colloquial term deepfakes (people don't usually want to be associated with the word "fake"), the technology is increasingly being used commercially. “More politely called AI-generated videos, or synthetic media, usage is growing rapidly in sectors including news, entertainment and education, with the technology becoming increasingly sophisticated.” “One of the early commercial adopters has been Synthesia, a London-based firm that creates AI-powered corporate training videos for the likes of global advertising firm WPP and business consultancy Accenture. ‘This is the future of content creation,’ says Synthesia chief executive and cofounder Victor Riparbelli.” |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
Jan21: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Machine Learning Prediction of Critical Transition and System Collapse”, by Kong et al | Complex dynamical systems, like climate, or an electrical power grid, have multiple cause and effect relationships, many of which operate in a time delayed and non-linear manner. These systems are characterized by “tipping points” (also, “phase transitions” or “critical transitions”) when they either shift from one regime to a very different one, or into a zone characterized by highly unstable or chaotic dynamics. The authors observe that, “to predict a critical transition… without relying on model is an outstanding problem in nonlinear dynamics and applied fields. A closely related problem is to predict whether a system is already in or if the system will be in a transient state preceding its collapse.” They develop a model free, machine learning-based solution to both problems that focuses on the evolution of key parameters to predict critical transitions. While this is an important development, it is not (yet) the same thing as being able to accurately predict critical transitions in complex adaptive systems (e.g., like the economy or financial markets) in which multiple intelligent agents are constantly using feedback about the impact of their behavior (and about other agents’ behaviors) to adapt their strategies to achieve goals that themselves may also be evolving over time. |
“Making Sense of Sensory Input”, by Evans, et al | SURPRISE Per Judea Pearl, current AI technologies have yet to meet two critical challenges: Causal and Counterfactual reasoning. Various researchers are coming at this fundamental problem from different directions. This paper provides further evidence of progress towards that goal. To be sure, there is still a lot of ground to cover between breakthroughs like this in labs, and their broad application in the economy. But recognizing these breakthroughs is critical to understanding the new AI capabilities that will be deployed in the future. The authors of this paper “attempt to answer a central question in unsupervised learning: what does it mean to “make sense” of a stream of sensory information? In [their] formalization, making sense involves constructing a symbolic causal theory that both explains the sensory sequence and also satisfies a set of unity conditions. The unity conditions insist that the constituents of the causal theory – objects, properties, and laws – must be integrated into a coherent whole” … “We believe there is more to “making sense” than prediction, retrodiction [explanation], and imputation [of missing values in a sequence]. Predicting the future state of [a system] may be part of what is involved in making sense – but it is not on its own sufficient. The ability to predict, retrodict, and impute is a sign, a surface manifestation, that one has made sense of the input. We want to define the underlying mental model that is constructed when one makes sense of the sensory input, and to show how constructing this mental model ipso facto enables one to predict, retrodict, and impute… They “assume that making sense of sensory input involves constructing a symbolic theory that explains the sensory input” … The authors’ second contribution is “a computer implementation, the ‘Apperception Engine’ that is designed to satisfy the above requirements. Our system is able to produce interpretable human-readable causal theories from very small amounts of data… A causal theory produced by our system is able to predict future sensor readings, as well as retrodict [explain] earlier readings, and impute (fill in the blanks of) missing sensory readings, in any combination. In fact, it is able to do all three tasks simultaneously.” |
“Understanding in Artificial Intelligence”, by Maetschke et al from IBM Research | Similar to the paper above, this one surveys the extent to which AI’s substantially improved predictive capabilities have been accompanied by increase in understanding (“to know why or how something happens or works”). Like others, the authors find that the answer is “much less.” They also review the strengths and shortcomings of methods using machine learning and symbolic reasoning, and discuss more promising approaches to true understanding. |
“Who is Winning the AI Race?” by Castro and McLaughlin, and “The Innovation Wars”, by Darby and Sewall | Our approach to macro forecasting acknowledges the immense challenges involved in predicting the effects produced by complex adaptive systems, which grow exponentially more difficult as the time horizon lengthens. However, we also stress that gaining a “coarse grained understanding” of the dynamics of such systems, along with the use of disciplined forecasting methods, can lead to predictions that are more accurate than chance. With respect to global macro (which is a system of underlying and interacting CAS like technology, national security, the economy, society, politics, etc.) a key aspect of this “coarse grained understanding” is recognition of a rough time sequence driving cause and effect (albeit with many feedback loops). This sequence begins with technology. Indeed, as Brian Arthur noted in his book “The Nature of Technology”, the economy is an expression of its technologies. Hence, it is critically important that we stay aware of developments in the areas of technology that are likely to have the largest effects as they develop and diffuse. The first of these papers analyzes the current state of competition in AI. It finds that while the US is ahead (based on the metrics the authors use), China is catching up quickly, while Europe is falling further behind. The second paper takes a broader look at shortcomings in the US approach to technology innovation. While they have been present for years, their negative effects have become much more visible as China has employed a very different approach (“civil-military fusion”) to make dramatic progress across multiple technologies. While the authors offer ideas for fixing the problems they identify in the US approach, many of these have been offered before, but have not been able to overcome the many political obstacles that have blocked their adoption in the past. Whether intensifying conflict with China will lead to a different outcome this time remains to be seen. |
“How Could Future AI Help Tackle Global Complex Problems?” by Anne-Marie Grisogono | SUPRRISE Anne-Marie Grisogono spent most of her career at Australia’s Defence Science and Technology Organization. Over the years, I have consistently found her publicly available research on the application of complex adaptive systems theory to practical problems to be among the best I’ve read. A few years ago she left for academia, and recently published this outstanding paper. In a world of increasing complexity (which Joseph Tainter’s research has hypothesized is a cause of civilization collapse), Grisogono focuses on the “wicked problems” created by rising complexity, and why only 10% of humans are good at solving them. She then very practically lays out what AI will need to do in the future in order to augment human beings’ ability to manage wicked problems, and ethical challenges such powerful AI will create. For all these reasons, it’s a great read. |
| |
| |
| |
| |
| |
| |
| |
| |
Dec20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Human Vs. Supervised Machine Learning: Who Learns Patterns Faster?” by Kuhl et al | We continue to focus on indicators and surprises related to the critical uncertainties surrounding rate at which AI technologies are being developed and deployed. This paper is closely related to another new analysis, on “Artificial Intelligence In War: Human Judgment as an Organizational Strength and a Strategic Liability”, by Goldfarb and Lindsay from the Brookings Institution. Their core contention is that as AI improves, human decision makers will be confronted with an exponentially large number of automated predictions they can choose from to inform and support their decision. Effectively, this is just taking “information overload” to a new level. The authors’ key claim is that this will now exponentially increase demand for people with good judgment skills, who can integration multiple AI-generated prediction when making critical decisions under time pressure. Kuhl et al complement this view with their finding that, when it comes to recognizing patterns (and making predictions based on them), when information is scarce, AI/machine learning still lags behind human beings. They note that, “supervised machine learning (SML), with its capabilities to support, or even replace, human workers in their daily tasks, is omnipresent in current discussions…tasks where machine learning models outperform humans are increasing… “But can this development be observed across all tasks? Current examples of (supervised) machine learning models outperforming humans are mainly present in areas where a high amount of training data is available, for example billions of played Go games or millions of labeled images… “In real life, however, often only limited training" data is available, sometimes just a single instance In this article, we are especially interested in learning patterns by humans and machines in data with only a few instances on a given task…“Empirical work in the comparison to human learning is still rare… The investigation of the learning curves, meaning the relation of required training samples and the resulting performance of humans in comparison to SML models, is a topic that has not yet been investigated.” The authors focus on this question: “How does the learning performance of humans and supervised machine learning models differ with limited training data?” They find that, “human performance shows two key characteristics across all four rules to be learned: High accuracy when labeling the first five instances (no accuracy below 60%, which outperforms the supervised machine learning models of three of the four rules) and only small performance improvements after learning with 20 or more training instances… “An explanation for the first observation is grounded in the concept of one-shot learning. Besides incremental learning, where humans learn step-by-step through trial and error, a human is also capable of one-shot learning, which is a technique to learn from a single instance. When a child touches a hot stove plate, he/she will immediately learn not to do it again… “Our interpretation of the data suggests that 20 training instances are the limit for humans' working memory, with more training instances only leading to cognitive overload and not to improved performance… “There are two general findings regarding the machine learning models across all four models: First, The performance after five training instances is similar or lower compared to the human performance and the machine learning models' performance correlates negatively with the complexity of the individual rules. “The second finding relates to one-shot learning. In contrast to humans, all three machine learning models can only do incremental learning. |
“On the Binding Problem in Artificial Neural Networks”, by Greff et al | “Existing neural networks fall short of human-level generalization. They require large amounts of data, struggle with transfer to novel tasks, and are fragile under distributional shift. “However, under the right conditions, they have shown a remarkable capacity for learning and modeling complex statistical structure in real-world data. One explanation for this discrepancy is that neural networks mostly learn about surface statistics in place of the underlying concepts, which prevents them from generalizing systematically. “However, despite considerable effort to address this issue, human-level generalization remains a major open problem. “In this paper, we will view the inability of contemporary neural networks to effectively form, represent, and relate symbol-like entities, as the root cause of this problem. “Contemporary neural networks still fall short of human-level generalization, which extends far beyond our direct experiences. In this paper, we argue that the underlying cause for this shortcoming is their inability to dynamically and flexibly bind information that is distributed throughout the network. “This "binding problem" affects their capacity to acquire a compositional understanding of the world in terms of symbol-like entities (like objects), which is crucial for generalizing in predictable and systematic ways. To address this issue, we propose a unifying framework that revolves around forming meaningful entities from unstructured sensory inputs (segregation), maintaining this separation of information at a representational level (representation), and using these entities to construct new inferences, predictions, and behaviors (composition). “Our analysis draws inspiration from a wealth of research in neuroscience and cognitive psychology, and surveys relevant mechanisms from the machine learning literature, to help identify a combination of inductive biases that allow symbolic information processing to emerge naturally in neural networks. We believe that a compositional approach to AI, in terms of grounded symbol-like representations, is o fundamental importance for realizing human-level generalization.” |
“A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications and Challenges”, by Adbar, et al | Uncertainty Quantification (UQ) is critically important in both optimization and decision processes. As Yakov Ben-Haim and Francois Hemez demonstrated in 2012, “Robustness to Uncertainty” is one of the three unavoidable dimensions that must be traded off in predictive modeling (along with Fidelity to Historical Data and Confidence Across Multiple Models). Unfortunately, many modelers don’t even realize these tradeoffs exist. (See Ben-Haim and Hemez’ paper, “Robustness, Fidelity and Prediction-Looseness of Models”). This new paper is very thorough overview how Uncertainty Quantification applies to Deep Learning. |
“Canaries in Technology Mines: Warning Signs of Transformative Progress in AI”, by Cremer and Whittlestone | SURPRISE “In this paper we introduce a methodology for identifying early warning signs of transformative progress in AI, to aid anticipatory governance and research prioritization… "We call these key milestones ‘canaries’ based on the colloquial phrase ‘canary in a coal mine’ to describe advance warning of an extreme event: in this case, advance warning of transformative AI. “We present results from an initial implementation to identify canaries for progress towards high-level machine intelligence (HLMI).” The authors highlight many warning signs and how they are related to each other. Many of these “canaries” are linked to the development of two critical capabilities: “Symbol-like representations: the ability to construct abstract, discrete and disentangled representations of inputs, to allow for efficiency and variable-binding. We hypothesise that this capability underpins several others, including grammar, mathematical reasoning, concept formation, and flexible memory. “Flexible memory: the ability to store, recognise, and re-use knowledge. We hypothesise that this ability would unlock many others, including the ability to learn from dynamic data, the ability to learn in a continual fashion, and the ability to learn how to learn.” |
The Second Montreal Debate on the Future of AI was held on December 23, 2020. It attracted multiple leaders who shared their views about the biggest challenges facing the field. | Highlights included the following: “Big data and deep learning alone won’t be enough to get to Artificial General Intelligence.” “The original and fundamental function of the human nervous system is to link perception to action. Intelligence emerges from active perception and interaction with the external environment. The next AI Northstar is how to apply evolutionary processes.” Multiple speakers agreed that causal and counterfactual reasoning are critically important and as yet unmet challenges. “We need to solve the binding problem” – see earlier Evidence Note. “Curiosity is still a big unsolved AI problem.” “How to reduce dependence on large training data and lots of compute to do deep learning?” “Transfer learning is still an unmet challenge.” “As a science, AI is still in the phenomenology and engineering phase; we have only begun the search for AI laws and physics.” |
“Advanced Technologies Adoption and Use By U.S. Firms: Evidence From The Annual Business Survey” by Zolas et al “AI and Jobs: Evidence from Online Vacancies”, by Acemoglu, Autor, et al “Artificial Intelligence In War: Human Judgment As An Organizational Strength And A Strategic Liability”, by Goldfarb and Lindsay from Brookings | Zolas et al find that, “advanced technology adoption is rare and generally skewed towards larger and older firms. Adoption patterns are consistent with a hierarchy of increasing technological sophistication, in which most firms that adopt AI or other advanced business technologies also use the other, more widely diffused technologies. Finally, while few firms are at the technology frontier, they tend to be large so technology exposure of the average worker is significantly higher.” Acemoglu, Autor et al “study the impact of AI on labor markets, using establishment level data on vacancies with detailed occupational information comprising the near-universe of online vacancies in the US from 2010 onwards. We classify establishments as "AI exposed" when their workers engage in tasks that are compatible with current AI capabilities.” They “document rapid growth in AI related vacancies over 2010-2018 that is not limited to the Professional and Business Services and Information Technology sectors and is significantly greater in AI-exposed establishments. AI exposed establishments are differentially eliminating vacancy postings that list a range of previously-posted skills while simultaneously posting skill requirements that were not previously listed. “Establishment-level estimates suggest that AI-exposed establishments are reducing hiring in non-AI positions as they expand AI hiring. However, we find no discernible relationship between AI exposure and employment or wage growth at the occupation or industry level, implying that AI is currently substituting for humans in a subset of tasks but it is not yet having detectable aggregate labor market consequences.” Goldfarb and Lindsay highlight the increasing importance (and growing shortage) of people with outstanding judgment skills as improvements in AI increase the volume of automated predictions they will confront when making decisions in increasingly high-velocity environments. |
“Capitol Hill — The 9/11 Moment Of Social Media” by Thierry Breton, European Commissioner for the Internal Market | SURPRISE The reaction of major tech companies to the January 6th attack on the US Capitol is likely to shift US opinion towards the emerging European approach to regulating the industry. “Just as 9/11 marked a paradigm shift for global security, 20 years later we are witnessing a before-and-after in the role of digital platforms in our democracy. “Social media companies have blocked U.S. President Donald Trump’s accounts on the grounds that his messages threatened democracy and incited hatred and violence. In doing so, they have recognized their responsibility, duty and means to prevent the spread of illegal viral content. They can no longer hide their responsibility toward society by arguing that they merely provide hosting services. “The dogma anchored in section 230 — the U.S. legislation that provides social media companies with immunity from civil liability for content posted by their users — has collapsed. “If there was anyone out there who still doubted that online platforms have become systemic actors in our societies and democracies, last week’s events on Capitol Hill is their answer. What happens online doesn’t just stay online: It has — and even exacerbates — consequences “in real life” too. “The unprecedented reactions of online platforms in response to the riots have left us wondering: Why did they fail to prevent the fake news and hate speech leading to the attack on Wednesday in the first place? “Regardless of whether silencing a standing president was the right thing to do, should that decision be in the hands of a tech company with no democratic legitimacy or oversight? Can these platforms still argue that they have no say over what their users are posting?” “Last week’s insurrection marked the culminating point of years of hate speech, incitement to violence, disinformation and destabilization strategies that were allowed to spread without restraint over well-known social networks. The unrest in Washington is proof that a powerful yet unregulated digital space — reminiscent of the Wild West — has a profound impact on the very foundations of our modern democracies. “The fact that a CEO can pull the plug on POTUS’s loudspeaker without any checks and balances is perplexing. It is not only confirmation of the power of these platforms, but it also displays deep weaknesses in the way our society is organized in the digital space.” |
| |
| |
| |
| |
| |
| |
Nov20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Smart Home Revolution Tests Legal Liability Regimes”, by Adam Green in the Financial Times | “When technology goes wrong — such as smart doorbells catching fire — the legal ramifications can be hazy. Many consumer products are not merely internet-connected but also have the ability to adapt, thanks to algorithms and machine learning. “’Autonomy and self-learning mean smart products are designed to evolve,’ says Rod Freeman, a product liability lawyer at law firm Cooley. The definition of a product is becoming fuzzier as layers of software are woven into devices, and it may be unclear who is at fault for an accident or failure, he says.” Anybody who has ever managed a high-risk project with multiple subcontractors (examples from my experience include large LNG projects and drilling complicated gas wells) knows that allocating responsibility for the liability associated with different risks is a very non-trivial issue. The potential obstacles this poses for deployment of the Internet-of-Things (IoT) is almost certainly underestimated by many companies and customers. |
“Inductive Biases for Deep Learning of Higher-Level Cognition” by Goyal and Bengio | The authors are two leading AI theorist/practitioners, who have written an important paper about challenges still to be met on the way to artificial general intelligence. They note, “deep learning brought remarkable progress but needs to be extended in qualitative and not just quantitative ways (larger datasets and more computing resources). We argue that having larger and more diverse datasets is important but insufficient… “We make the case that evolutionary forces, the interactions between multiple agents, the non-stationary and competition systems put pressure on the learner to achieve the kind of flexibility, robustness and ability to adapt quickly which humans seem to have when they are faced with new environments. “In addition to thinking about the learning advantage, this paper focuses on knowledge representation in neural networks, with the idea that by decomposing knowledge in small pieces which can be recomposed dynamically as needed (to reason, imagine or explain at an explicit level), one may achieve the kind of systematic generalization which humans enjoy and is obvious in natural language.” |
“What AI Can Do for Football [Soccer] and What Football Can Do for AI”, by Tuyls et al from DeepMind | Many people lack a familiar context within which they can understand the profound impact that artificial intelligence as a general purpose technology will almost certainly have in the coming years. This paper provides that context (sports) and shows how the application of multiple AI and other advanced technologies are already having a profound impact. |
“Underspecification Presents Challenges for Credibility in Modern Machine Learning”, by D’Amour et al, and “A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications, and Challenges”, by Abdar et al | It is one thing to develop and deploy AI tools; it is another for human beings to trust them in actual use. These two papers approach the trust in AI issue by opening up the black box, so to speak. D’Amour and his colleagues highlight the sources of imperfect categorization and prediction in current machine learning models. Abdar and his colleagues focus on a critical issue related to human trust in AI models – the extent to which they quantify the degree (and ideally the sources) of uncertainty associated with their results. |
| |
| |
| |
| |
| |
| |
| |
| |
| |
Oct20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Why Modeling the Spread of COVID is So Damn Hard” by Matthew Hutson | The author begins by acknowledging the elephant in the room: “Too many of the COVID-19 models have led policy makers astray.” He then presents a clear explanation of the shortcomings of two common modeling approaches. “A SEIRS model puts people into categories: susceptible (S), exposed (E), infected (I), removed from the susceptible population (R), and potentially back to susceptible (S) again, depending on whether a recovered person has immunity from the disease. The modeler’s job is to define the equations that determine how people move from one category to the next. Those equations depend on a wide variety of parameters drawn from biology, behavior, politics, the economy, the weather, and more.” Unfortunately, this system is evolving over time. If the parameters aren’t accurately updated at the same speed the system is evolving, the model becomes progressively less accurate. Machine learning doesn’t start with a causal model; rather it starts with a large amount of training data for multiple variables, and a set of target variables to predict, and then uses techniques like neural networks to maximize predictive accuracy. However, these models suffer from three shortcomings: (1) a lack of training data about pandemics; (2) a complex system that is constantly evolving; and (3) the lack of a clear causal explanation of how the model predicted the future results from the input data. Hutson is most encouraged by agent-based models. “These are much like the video game The Sims. Each individual in a population is represented by their own bit of code, called an agent, which interacts with other agents as it moves around the world. “One of the most successful agent-based models was designed at the University of Sydney. The model has three layers, beginning with a layer of demographics. ‘We’re essentially creating a digital twin for every person represented in the census,’ said Mikhail Prokopenko, a computer scientist at the university. He and his colleagues built a virtual Australia comprising 24 million agents, whose distribution matches the real thing in terms of age, household size, neighborhood size, school size, and so on. “The second layer is mobility, in which agents are assigned to both a household and a school or workplace. On top of demographics and mobility, they add the [third layer] disease, including transmission rates within households, schools, and workplaces, and how the disease progresses in individuals… “When set in motion, the model ticks twice a day: People come in contact at school or work in the daytime, then at home at night. It’s like throwing dice over and over. The model covers 180 days in a few hours. The team typically runs tens or hundreds of copies of the model in parallel on a computing cluster to generate a range of outcomes. “The biggest insight was that social distancing helps very little if only 70 percent of people practice it, but successfully squashes COVID-19 incidence if 80 percent of people can manage it over a span of a few months…Agent-based models look ideal for simulating possible interventions to guide policy, but they’re a lot of work to build and tricky to calibrate.” |
“Heterogeneous Multi-Agent Reinforcement Learning for Unknown Environment Mapping”, by Wakilpoor et al | This paper reports important progress in combining reinforcement learning (a common machine learning method) with agent based modeling, to automatically develop situation awareness in an evolving complex environment. Specifically, unmanned aerial surveillance platforms constantly collect and share data about the evolving environment, and use a novel reinforcement learning algorithm to update their collective understanding of the situation and then use it to adjust their behavior to accelerate their learning. The authors note that from these agent interactions more efficient, cooperative behaviors emerge, without human intervention. |
“Deep Generative Modeling in Network Science with Applications to Public Policy Research”, by Hartnett et al from RAND “Network data is increasingly being used in quantitative, data-driven public policy research. These are typically very rich datasets that contain complex correlations and inter-dependencies. This richness both promises to be quite useful for policy research, while at the same time posing a challenge for the useful extraction of information from these datasets - a challenge which calls for new data analysis methods. In this report, we formulate a research agenda of key methodological problems whose solutions would enable new advances across many areas of policy research. We then review recent advances in applying deep learning to network data, and show how these methods may be used to address many of the methodological problems we identified. We particularly emphasize deep generative methods, which can be used to generate realistic synthetic networks useful for microsimulation and agent-based models capable of informing key public policy questions.” | It is increasingly clear that improvements in agent based modeling technologies, especially when they are combined with machine learning, are very likely to have a substantial impact on our ability to understand and predict the behavior of complex adaptive systems. This new paper is an excellent overview of the significant progress that has been made in this area in recent years, as well as the critical obstacles that have not yet been overcome and key indicators to monitor. |
“Less Than One-Shot Learning: Learning N Classes from M “Deep supervised learning models are extremely data-hungry, generally requiring a very large number of samples to train on. Meanwhile, it appears that humans can quickly generalize from a tiny number of examples. “Getting machines to learn from ‘small’ data is an important aspect of trying to bridge this gap in abilities. “Few-shot learning (FSL) is one approach to making models more sample-efficient. In this setting, models must learn to discern new classes given only a few examples. “Further progress in this area has enabled a more extreme form of FSL called one-shot learning (OSL); a difficult task where models must learn to discern a new class given only a single example of it. “In this paper, we propose ‘less than one’-shot learning (LO-shot learning), a setting where a model must learn N new classes given onlyM < N examples, less than one example per class… “We show that this is achievable.” | SURPRISE Today, improvements in both the capabilities of machine learning technologies and their application are held back by three broad constraints. First, they have required large amounts of training data. Second, they have been based on associative reasoning, rather than more powerful causal and counterfactual reasoning. Third, they have required large amounts of computing power (“compute”). This paper shows that the first constraint is on its way to becoming much less binding. |
“Algorithms for Causal Reasoning in Probability Trees”, by Genewein et al from DeepMind | SURPRISE This paper is an example of progress towards relaxing the second constraint noted above. “A probability tree is one of the simplest models for representing the causal generative process of a random experiment or stochastic process The semantics are self-explanatory: each node in the tree corresponds to a potential state of the process, and the arrows indicate both the probabilistic transitions and the causal dependencies between them… “Our work is the first to provide concrete algorithms for (a) computing minimal representations of arbitrary events formed through propositional calculus and causal precedences; and (b) computing the three fundamental operations of the causal hierarchy, namely conditions, interventions, and counterfactuals.” |
| |
| |
| |
| |
| |
| |
| |
| |
Sep20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Investigation of Competition in Digital Markets”, by the Subcommittee on Antitrust, of the US House of Representatives See also, “Concentrated Power of Big Tech Harms the US” by Rana Foroohar in the Financial Times | This long awaited report found multiple instances of large technology leaders exercising monopoly power. If Democrats end up controlling the White House and both the US House and Senate after the November election, this report will likely form the basis of much more aggressive government action against these companies. This will also likely lead to closer cooperation between the US and EU on the regulation of technology companies. From the Report: “On June 3, 2019, the House Judiciary Committee announced a bipartisan investigation into competition in digital markets,2 led by the Subcommittee on Antitrust, Commercial, and Administrative Law. "The purpose of the investigation was to (1) document competition problems in digital markets; (2) examine whether dominant firms are engaging in anticompetitive conduct; and (3) assess whether existing antitrust laws, competition policies, and current enforcement levels are adequate to address these issues” … “Over the past decade, the digital economy has become highly concentrated and prone to monopolization. Several markets investigated by the Subcommittee—such as social networking, general online search, and online advertising—are dominated by just one or two firms. The companies investigated by the Subcommittee—Amazon, Apple, Facebook, and Google—have captured control over key channels of distribution and have come to function as gatekeepers. "Just a decade into the future, 30% of the world’s gross economic output may lie with these firms, and just a handful of others. “In interviews with Subcommittee staff, numerous businesses described how dominant platforms exploit their gatekeeper power to dictate terms and extract concessions that no one would reasonably consent to in a competitive market” … “This significant and durable market power is due to several factors, including a high volume of acquisitions by the dominant platforms. Together, the firms investigated by the Subcommittee have acquired hundreds of companies just in the last ten years. In some cases, a dominant firm evidently acquired nascent or potential competitors to neutralize a competitive threat or to maintain and expand the firm’s dominance. "In other cases, a dominant firm acquired smaller companies to shut them down or discontinue underlying products entirely—transactions aptly described as “killer acquisitions.” “In the overwhelming number of cases, the antitrust agencies did not request additional information and documentary material under their pre-merger review authority in the Clayton Act, to examine whether the proposed acquisition may substantially lessen competition or tend to create a monopoly if allowed to proceed as proposed. For example, of Facebook’s nearly 100 acquisitions, the Federal Trade Commission engaged in an extensive investigation of just one acquisition: Facebook’s purchase of Instagram in 2012. “During the investigation, Subcommittee staff found evidence of monopolization and monopoly power.” As the Financial Times noted, “The best way to police [large tech companies] is to reinvent a US antitrust model still based on the concept of consumer harm, or whether prices are being driven up. The focus should be broadened to the impact of corporate power on market structure, competition, innovation, and quality (“The Growing Momentum Towards Curbing Big Tech”). |
Far less remarked upon than the antitrust report was a new proposal by the Trump administration to repeal Section 230 of the 1996 Communications Decency Act – and Joe Biden’s apparent agreement with it. | In the early days of the internet, there was increasing litigation against technology companies that allowed community content creation for the nature of the content on their websites. This prompted the Communications Decency Act, in which section 230 states that, unlike traditional media companies, “No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider”. This effectively eliminated the threat of litigation against technology companies operating network platforms (e.g., Facebook, Twitter, etc.). We’ve come a long way since then, and today, faced with allegations of foreign interference in American elections and controversies over “fake news”, platform companies have begun to censor different types of content. Predictably, this has led to protests from across the political spectrum. While antitrust is one approach to reducing the power of technology companies, so too is elimination or reform of Section 230, which would almost certainly result in very substantial changes to social and political dynamics in many countries. |
“From AI To Facial Recognition: How China Is Setting The Rules in New Tech”, by James Kynge and Nian Liu in the Financial Times | The authors note that, “the intensifying geopolitical battleground of technological standards, a much overlooked yet crucial aspect of a new struggle for global influence between China and the US. “Such standards might seem obscure, but they are a crucial element of modern technology. If the cold war was dominated by a race to build the most nuclear weapons, the contest between the US and China — as well as the EU — will partly be played out through a struggle to control the bureaucratic rule-setting that lies behind the most important industries of the age”… “Standard-setting has for decades largely been the preserve of a small group of industrialised democracies…But China now has other ideas”… “An intensifying US-China battle to dominate standards, especially in emerging technologies, could start to divide the world into different industrial blocs… Strategic competition between the US and China raises the spectre of a fragmentation of standards that creates a new technological divide.” |
“The Security of 5G”, by the Defence Committee of the UK House of Commons. | “Our inquiry into the security of 5G was launched in the context of a lively debate on the security of the UK’s 5G network in Parliament and across the country from late 2019 and through 2020 with a focus on the presence in our network of high-risk vendors, particularly Huawei… “The presence of Huawei equipment in our network increased the risk posed by cyber attacks and there is no doubt that Huawei’s designation as a high-risk vendor was justified. The Huawei Cyber Security Evaluation Centre consistently reported on its low-quality products and concerning approach to software development, which has resulted in increased risk to UK operators and networks. The presence of Huawei in the UK’s 5G networks posed a significant security risk to individuals and to our Government” … “A further geopolitical consideration our inquiry highlighted was Huawei’s relationship with the Chinese state. It is clearly strongly linked to the Chinese state and the Chinese Communist Party, despite its statements to the contrary, as evidenced by its ownership model and the subsidies it has received. Additionally, Huawei’s apparent willingness to support China’s intelligence agencies and China’s 2017 National Intelligence Law are further cause for concern. "Having a company so closely tied to a state and political organisation sometimes at odds with UK interests should be a point of concern and the decision to remove Huawei from our networks is further supported by these links. Concern about Huawei is based on clear evidence of collusion between the company and the Chinese Communist Party apparatus” … “China dominates the telecommunications industry and it is evident that the UK has a lack of industrial capacity in this sector. This is not unique to the UK and in order to combat China’s dominance, we support the principle of proposals for forming a D10 alliance of democracies to provide alternatives to Chinese technology.” |
“Multi-agent Social Reinforcement Learning Improves Generalization”, by Ndouse et al. “Social learning is a key component of human and animal intelligence. By taking cues from the behavior of experts in their environment, social learners can acquire sophisticated behavior and rapidly adapt to new circumstances. “This paper investigates whether independent reinforcement learning (RL) agents in a multi-agent environment can use social learning to improve their performance using cues from other agents… "We are able to train agents to leverage cues from experts to solve hard exploration tasks. The generalized social learning policy learned by these agents allows them to not only outperform the experts with which they trained, but also achieve better zero-shot transfer performance than solo learners when deployed to novel environments with experts”. “Importance Weighted Policy Learning and Adaption”, by Galashov et al from Google. The ability to exploit prior experience to solve novel problems rapidly is a hallmark of biological learning systems and of great practical importance for artificial ones. In the meta reinforcement learning literature much recent work has focused on the problem of optimizing the learning process itself. In this paper we study a complementary approach [that] achieves competitive adaptation performance compared to meta reinforcement learning baselines and can scale to complex sparse-reward scenarios.” “A Survey of Deep Meta Learning”, by Huisman et al. “Deep neural networks can achieve great successes when presented with large data sets and sufficient computational resources. However, their ability to learn new concepts quickly is quite limited. Meta-learning is one approach to address this issue, by enabling the network to learn how to learn. The exciting field of Deep Meta-Learning advances at great speed, but lacks a unified, insightful overview of current techniques. This work presents just that.” | SURPRISE These new papers provide evidence of further progress in critical areas of machine learning and artificial intelligence. Today, artificial intelligence is almost always based on what Judea Pearl refers to a “associative” reasoning (e.g., complex statistical relationships and pattern matching). It has not yet been able to reach Pearl’s higher levels of causal and counterfactual reasoning (particularly involving complex adaptive systems and networks), which will require continued progress in many technologies (see Pearl’s “The Book of Why”). When and if artificial intelligence is able to incorporate causal and counterfactual reasoning, it will very likely result in an order of magnitude (or more) improvement in its performance and range of applications. For this reason, we continuously look for indicators of progress in this area. |
“How to Talk When a Machine is Listening: Corporate Disclosure in the Age of AI”, by Cao et al “This paper analyzes how corporate disclosure has been reshaped by machine processors, employed by algorithmic traders, robot investment advisors, and quantitative analysts. Our findings indicate that increasing machine and AI readership, proxied by machine downloads, motivates firms to prepare filings that are friendlier to machine parsing and processing. “Moreover, firms with high expected machine downloads manage textual sentiment and audio emotion in ways catered to machine and AI readers, such as by differentially avoiding words that are perceived as negative by computational algorithms as compared to those by human readers, and by exhibiting speech emotion favored by machine learning software processors.” | This paper is an excellent example of a critical challenge faced by the deployment of artificial intelligence technologies in systems (and especially social systems) where agents can adapt to their presence and in so doing degrade their expected performance. Another example is algorithm versus algorithm trading in financial markets. See also this month’s National Security Evidence File, on how automated cycles of rapid adaptation can lead to uncontrolled escalation in the realm of cyberwar. |
| |
| |
| |
| |
| |
| |
| |
Aug20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
At the end of July, the US Senate passed the FY 2021 National Defense Authorization Act (NDAA). The House had previously passed its version of the bill. Both contain billions of dollars in tax credits and grants for construction of new semiconductor manufacturing facilities in the United States. | In May, the US prohibited (on national security grounds) the use of exported semiconductor manufacturing equipment to produce chips for use by a group of foreign owned companies, including Huawei. In our June issue, we detailed how this could significantly increase US-China tensions, because TSMC, the world’s leading semiconductor manufacturing company has most of its manufacturing capacity located in Taiwan. At the end of July, this became even more concerning when Intel, which as significant US-based semiconductor manufacturing capacity in the United States, announced a six month delay in the start of production for its 7 nanometer chip. This signified the widening of its technology gap with TSMC, which is already manufacturing a 5nm chip (which have about the same processor density as Intel’s 7nm chip), and is scheduled to begin producing a 3nm chip in late 2022 (see, “ TSMC Keeps Pushing the Envelope -- That's Good News for Apple, AMD and Others. If TSMC makes good on its latest promise, it could hold onto its recently-won chip manufacturing lead for some time”, by Eric Johnsa) This apparently led to Intel’s announcement that it might consider outsourcing more chip production. This would, potentially further raise the strategic importance (in both economic and national security terms) of TSMC’s Taiwan plants. Hence the importance of additional funding in the NDAA for domestic US chip production. |
In August, an AI pilot defeated a top US Air Force F-16 pilot in a simulated dogfight, as part of a DARPA’s Air Combat Evolution (ACE) program. | This was an important step towards the goal of eventually pairing a number of drones with a human pilot (a concept which also applies on land and on and under the sea). On its own, this will substantially change the character of conflict. However, it also highlights questions related to the differences between automated and autonomous weapons systems. As former Navy pilot and current professor Missing Cummings describes in “Lethal Autonomous Weapons: Meaningful Human Control or Meaningful Human Certification?”, “it is important to define autonomy in technology, which is not the same as automation. “Automated systems operate by clear repeatable rules based on unambiguous sensed data. “Autonomous systems take in data about the unstructured world around them, process that data to generate information, and generate alternatives and make decisions in the face of uncertainty. Often, people will refer to this set of capabilities as self-governing. Systems are not necessarily either fully automated or fully autonomous, but often fall somewhere in between… “Currently autonomous technologies perform best when they are doing a very narrow task with little to no uncertainty in their environments…The limitations of machine learning and computer vision systems are currently the Achilles’ heel for all autonomous systems, military and civilian.” Finally, the United States is not alone in developing lethal autonomous weapons systems. August also saw the publication of speculations that a new amphibious assault ship being built by the Chinese Navy will carry a significant number of unmanned combat aerial vehicles. |
Around the world, reopening of schools is underway (with varying degrees of chaos) using a range of remote, hybrid, and in-person models. A critical issue that in the US many districts are avoiding is the extent of learning losses students suffered following last spring’s haphazard shift to remote education, and whether they will ever be recovered. | As I wrote in an analysis published by the Fordham Institute, analysis by NWEA has found that learning losses have been extensive. Previous research by ACT, Inc. provides base rate evidence that recovery of learning losses is very difficult for most students, given current school system configurations. Given scarce talent issues that many companies were already struggling with before COVID arrived, failure to recover student learning losses has negative implications for future productivity and economic growth, and thus for inequality and social and political conflict. |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
Jul20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
In July, OpenAI released GPT-3, the latest version of its natural language processing model, which has substantially improved performance over GPT-2 that was released last year. NLP is the subset of AI technologies that focuses on analyzing and generating human language – e.g., to answer questions posed by humans. | SURPRISE GPT-3 will enable many more creative and potentially disruptive applications of NLP. However, it is also important to keep in mind that GPT-3 is still based on associative, not causal or counterfactual reasoning. It excels at matching patterns (e.g., to answer questions or complete sentences), but does not actually comprehend the meaning of language or integrated it with existing knowledge to build and apply mental models of the real world in the way humans do. |
New surveys from McKinsey and Deloitte on the state of AI adoption by businesses after COVID reached very similar conclusions | SURPRISE In “The Productivity J-Curve”, Brynjolfsson, Rock, and Syverson concluded that (as has been the case with previous general purpose technology innovations) substantial improvements in AI will only produce large productivity gains once companies have made substantial investments in intangibles, like improved employee knowledge and skills, and reconfigured processes and organizational structures. These two new reports are important indicators of the speed at which these changes are happening on the ground. In “Entering a New Decade of AI: The State of Play”, McKinsey finds that while growth in AI adoption is accelerating, “less than a third of companies that we surveyed have deployed AI in multiple businesses or functions.” The report highlights a familiar reason for this: “it’s hard work to deploy technology within an organization, not only because the technology problems are hard, but also because the change management is really hard.” McKinsey also noted that “frontier technologies” have not yet seen wide deployments. These include reinforcement learning and Generative Adversarial Networks (GANs). Also of note was McKinsey’s finding a significant gap between companies’ recognition of various AI related risks and their perceived ability to manage them today: “we asked respondents to identify which risks are relevant and then which risks they have mitigated. And across a certain set of risks—such as cybersecurity, explainability, regulatory compliance, and others—among all respondents we still see a significant delta between respondents who say their company has identified a relevant risk but then have successfully been able to mitigate it.” Finally, McKinsey found “more respondents than in the past year saying that there potentially could be a decrease in the size of their workforce for an individual company as a result of the deployment of AI.” In “Thriving in the Era of Pervasive AI”, Deloitte’s survey finds that “virtually all adopters are using AI to improve efficiency; mature adopters are also harnessing the technologies to boost differentiation.” However, “more respondents than in the past year saying that there potentially could be a decrease in the size of their workforce for an individual company as a result of the deployment of AI.” Like McKinsey, Deloitte also found a large gap between companies recognition of AI related risks and their perceived ability to manage them: “more respondents than in the past year saying that there potentially could be a decrease in the size of their workforce for an individual company as a result of the deployment of AI.” For example, while 54% of respondents cited “Making Bad Decisions Based on AI’s Recommendations” as a “major/extreme concern”, only 38% said they were prepared to manage it. Finally, 53% of respondents cited job losses related to AI as a “major/extreme concern”, but only 37% believed they were prepared to manage it. |
“The Deck is Not Stacked: Poker and the Limits of AI”, Maria Konnikova | SURPRISE Konnikova begins by noting that, “the great game theorist John von Neumann viewed poker as the perfect model for human decision making, for finding the balance between skill and chance that accompanies our ever choice. He saw poker as the ultimate strategic challenge, combining as it does not just the mathematical elements of a game like chess, but the uniquely human, psychological angles that are more difficult to model precisely.” She goes on to describe the research of Carnegie Mellon professor Tuomas Sandholm, whose team has designed the two most effective poker playing AI’s, Libratus and Pluribus. Sandholm notes that his “goal isn’t to solve poker, as such, but to create algorithms whose decision making prowess in poker’s world of imperfect information and stochastic situations — situations that are randomly determined and unable to be predicted — can then be applied to other stochastic realms, like the military, business, government, cybersecurity, even health care.” Last year, Pluribus became the first AI to defeat multiple top ranked players (see, “Superhuman AI for Multiplayer Poker”, by Brown and Sandholm) Two of the Pluribus team’s innovations stand out. The first was its use of regret minimization as its objective function – e.g., how much better were the results from actions not chosen than the one that was. In humans, regret triggers powerful emotions that cloud our ability to learn from it. AI therefore learns faster from regret (see, “Stochastic Regret Minimization in Extensive Form Games” by Farina et al). The second was the development of an AI that could learn effective strategies by observing a game, without knowing its rules (see, “Efficient Exploration of Zero-Sum Stochastic Games” by Martin et al). While Pluribus’ achievement is extremely impressive, Konnikova concludes with a critical distinction: “real-life applications have to contend with something that a poker algorithm does not: the weights that are assigned to different elements of a decision” [e.g., to multiple competing goals]. If different players are pursuing different goals, or put different weights on them, and/or on different causal theories in a complex adaptive system (or both, as in “wicked problems”), then AI techniques are still insufficient. But as Pluribus shows, they are quickly improving. |
“Standardising the Splinternet: How China’s technical standards could fragment the Internet”, by Hoffman et al | SURPRISE “China’s drive for technological dominance has resulted in a long-term, government-driven national strategy. This includes the creation of native technologies which reflect local policies and politics, micromanagement of the Internet from the top down, and the use of international standards development organisations (SDOs), such as the UN agency the International Telecommunication Union (ITU), to legitimize and protect these technologies in the global marketplace. “Alternate Internet technologies based on a new ‘decentralized Internet infrastructure’ are being developed in SDOs and marketed by Chinese companies. In a worst-case scenario, these alternate technologies and a suite of supporting standards could splinter the global Internet’s shared and ubiquitous architecture… “A fragmented network would introduce new challenges to cyber defence and could provide adversaries with a technical means to undermine norms, predictability and security of today’s cyberspace – which would also impact human rights and widen the digital divide.” |
“Event Prediction in Big Data Era: A Systematic Survey” by Liang Zhao from Emory University | “Events are occurrences in specific locations, time, and semantics that nontrivially impact either our society or the nature, such as earthquakes, civil unrest, system failures, pandemics, and crimes. It is highly desirable to be able to anticipate the occurrence of such events in advance in order to reduce the potential social upheaval and damage caused. “Event prediction, which has traditionally been prohibitively challenging, is now becoming a viable option in the big data era and is thus experiencing rapid growth, also thanks to advances in high performance computers and new Artificial Intelligence techniques… “Due to the strong interdisciplinary nature of event prediction problems, most existing event prediction methods were initially designed to deal with specific application domains, though the techniques and evaluation procedures utilized are usually generalizable across different domains. However, it is imperative yet difficult to cross-reference the techniques across different domains, given the absence of a comprehensive literature survey for event prediction. "This paper aims to provide a systematic and comprehensive survey of the technologies, applications, and evaluations of event prediction in the big data era.” After providing a thorough overview, Zhang addresses remaining challenges in event prediction, including one that occurs across multiple AI applications: Judea Pearl’s familiar distinction between associational (e.g., correlation), causal, and counterfactual reasoning. “The ultimate purpose of event prediction is usually not just to anticipate the future, but to change it, for example by avoiding a system failure and flattening the curve of a disease outbreak. However, it is difficult for practitioners to determine how to act appropriately and implement effective policies [actions] in order to achieve the desired results in the future. This requires a capability that goes beyond simply predicting future events based on the current situation, requiring them instead to also take into account the new actions being taken in real time and then predict how they might influence the future. “One promising direction is the use of counterfactual event prediction that models what would have happened if different circumstances had occurred. Another related direction is prescriptive analysis where different actions can be merged into the prediction system and future results anticipated or optimized. Related works have been developed in few domains such as epidemiology. However, as yet these lack sufficient research in many other domains that will be needed if we are to develop generic frameworks that can benefit different domains.” |
“Discovering Reinforcement Learning Algorithms”, by Oh et al from DeepMind and “MISIM: An End-to-End Neural Code Similarity System” by Ye et al from Intel Labs | SURPRISE These two papers both provide evidence of important advances in the automation of machine learning software development (known as “machine programming”). The DeepMind team notes that, “reinforcement learning (RL) has a clear objective: to maximise expected cumulative rewards (or average rewards), which is simple, yet general enough to capture many aspects of intelligence. Even though the objective of RL is simple, developing efficient algorithms to optimise such objective typically involves a tremendous research effort, from building theories to empirical investigations. “An appealing alternative approach is to automatically discover RL algorithms from data generated by interaction with a set of environments, which can be formulated as a meta-learning problem”… “Although there have been prior attempts at addressing this significant scientific challenge, it remains an open question whether it is feasible to discover alternatives to fundamental concepts of RL such as value functions and temporal-difference learning. “This paper introduces a new meta-learning approach that discovers an entire update rule which includes both ‘what to predict’ (e.g. value functions) and ‘how to learn from it’ (e.g. bootstrapping) by interacting with a set of environments.” The team from Intel notes that one challenge in machine programming has been “construction of accurate code similarity systems, which generally try to determine if two code snippets are semantically similar (i.e., having similar characteristics through some analysis). Accurate code similarity systems may assist in many programming tasks, such as code recommendation systems to improve programming development productivity, to automated bug detection and mitigation systems to improve programmer debugging productivity.” Their paper describes a substantial advance Intel has made in this area. |
In the United States, the reopening of K-12 schools has become another polarized political conflict, with teachers unions issuing a long list of demands (e.g., no new COVID cases for 14 days) before they return to the classrooms. Unfortunately, there is little evidence of improvement in the weak approaches to remote learning that parents and students experienced in the spring. | Education is a critical “social technology.” In the absence of strong education system performance, talent shortages constrain the diffusion of new technologies. This contributes to winner take all dynamics in different industries, as well as to income inequality, as firms that attract the talent needed to adopt new technologies earn higher profits and pay employees more. However, slowed technology diffusion holds down the overall rate of productivity improvement and economic growth in the economy, which has negative implications for future levels of social and political conflict. The sudden shift to remote learning that COVID forced on schools resulted in substantial learning losses, which look likely to worse. Unfortunately, base rate data indicate that, in the absence of substantial change to the US education system, it is very unlikely that these learning losses will be made up (e.g., see “Catching Up to College and Career Readiness”, by ACT Inc., and “COVID-19 And Student Learning In The United States: The Hurt Could Last A Lifetime”, by McKinsey). |
| |
| |
| |
| |
| |
| |
Jun20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Maintaining the Competitive Advantage in Artificial Intelligence and Machine Learning”, by Waltzman et al from RAND | SURPRISE “Artificial intelligence (AI) technologies hold the potential to become critical force multipliers in future armed conflicts. The People's Republic of China has identified AI as key to its goal of enhancing its national competitiveness and protecting its national security. “If its current AI plan is successful, China will achieve a substantial military advantage over the United States and its allies. That has significant negative strategic implications for the United States… Key conclusions: “It is difficult, perhaps impossible, to arrive at a definitive statement about which country has the lead in AI. It is more useful to talk about various parts of the AI ecosystem. It appears possible that the United States has a narrow lead in several key areas of AI, although China has several advantages and a high degree of leadership focus on this issue. “As of early 2020, the United States has a modest lead in AI technology development because of its substantial advantage in the advanced semiconductor sector. China is attempting to erode this edge through massive government investment. The lack of a substantial U.S. industrial policy also works to Chinese advantage. “China has an advantage over the United States in the area of big data sets that are essential to the development of AI applications. This is partly because data collection by the Chinese government and large Chinese tech companies is not constrained by privacy laws and protections. However, the Chinese advantage in data volume is probably insufficient to overcome the U.S. edge in semiconductors. “Breakthrough fundamental research is not a critical dimension for comparing U.S.-China relative competitive standing from a DoD perspective. Fundamental research, regardless of whether it is U.S., Chinese, or a U.S.-Chinese collaboration, is available to all. “Commercial industry is also not a critical dimension for competitive comparison. Industries with corporate headquarters in the United States and in China seek to provide products and services wherever the market is.” |
“Shaping the Terrain of AI Competition” by Tim Hwang, Georgetown University Center for Security and Emerging Technology | SURPRISE “Concern that China is well-positioned to overtake current U.S. leadership in artificial intelligence in the coming years has prompted a simply stated but challenging question. How should democracies effectively compete against authoritarian regimes in the AI space? “Policy researchers and defense strategists have offered possible paths forward in recent years, but the task is not an easy one. “Particularly challenging is the possibility that authoritarian regimes may possess structural advantages over liberal democracies in researching, designing, and deploying AI systems. Authoritarian states may enjoy easier access to data and an ability to coerce adoption of technologies that their democratic competitors lack.” “Authoritarians may also have stronger incentives to prioritize investments in machine learning, as the technology may significantly enhance surveillance and social control.” “No policy consensus has emerged on how the United States and other democracies can overcome these burdens without sacrificing their commitments to rights, accountability, and public participation.” “This paper offers one answer to this unsettled question in the form of a “terrain strategy.” It argues that the United States should leverage the malleability of the AI field and shape the direction of the technology to provide structural advantages to itself and other democracies. “This effort involves accelerating the development of certain areas within machine learning (ML)—the core technology driving the most dramatic advances I AI—to alter the global playing field… “Democracies should invest their resources in three critical domains: (1) Reducing a dependence on data. Authoritarian regimes may have structural advantages in marshaling data for ML applications when compared to their liberal democratic adversaries. To ensure better competitive parity, democracies should invest in techniques that reduce the scale of real-world data needed for training effective ML systems. (2) Fostering techniques that support democratic legitimacy. Democracies may face greater friction in deploying ML systems relative to authoritarian regimes due to their commitments to public consent. Enhancing the viability of the technology in a democratic society will require investing in ML subfields, including work in interpretability, fairness, and privacy. (3) Challenging the social control uses of ML. Recent advances in AI appeal to authoritarian regimes in part because they promise to enhance surveillance and other mechanisms of control. Democracies should advance research eroding the usefulness of these applications.” |
“The Evolutionary Dynamics of Independent Learning Agents in Population Games”, by Hu et al | SURPRISE The combination of reinforcement learning methods with simulations involving multiple adaptive agents has the potential to produce breakthroughs in our ability to understand and influence the outcomes produced by the complex adaptive systems that characterize modern life. Hence it is important to monitor indicators of progress in this critical area. The authors begin by noting that “understanding the evolutionary dynamics of reinforcement learning under multiagent settings has long remained an open problem.” (For an excellent previous review of this issue, see, “A Survey of Learning in Multiagent Environments: Dealing with Non-Stationarity”, by Hernanez-Leal et al). Most commonly, the behavior of these systems is modeled at the aggregate level, using stochastic equations in a top-down approach. Rather than focus on the more commonly studied 2-player (agent) games, Hu and her co-authors “ consider population games, which model the strategic interactions of a large population comprising small and anonymous agents.” Uniquely, they present a formal relationship between the top-down stochastic process of the system as a whole and the “dynamics of independent learning agents who reason based on reward signals.” |
“Empirically Verifying Hypotheses Using Reinforcement Learning”, by Marino et al | SURPRISE We have repeatedly noted the importance of indications that artificial intelligence technologies are moving beyond associational (e.g., correlational) methods, to higher levels of causal and counterfactual reasoning (e.g. as described by Judea Pearl in The Book of Why). While not employing Pearl’s causal methods, this paper is interesting because it attempts formulate hypothesis verification as a reinforcement learning problem (for a broader survey, see, “A Survey of Learning Causality with Data: Problems and Methods”, by Guo et al). The authors of this paper note that, “Empirical research on early learning has shown that infants build an understanding of the world around by constantly formulating hypotheses about how some physical aspect of the world might work and then proving or disproving them through deliberate play. “Through this process the child builds up a consistent causal understanding of the world. This contrasts with manner in which current ML systems operate… Learning settings use a single user-specified objective function that codifies a high-level task, and the optimization routine finds the set of parameters (weights) that maximizes performance on the task. "The learned representation (knowledge of how the world works) is embedded in the weights of the model - which makes it harder to inspect, hypothesize or even enforce domain constraints that might exist” … The authors “aim to build an agent that, given a hypothesis about the dynamics of the world, can take actions to generate observations which can help predict whether the hypothesis is true or false. Existing RL algorithms fail to solve this task, even for simple environments. In order to train the agents, we exploit the underlying structure of many hypotheses, factorizing them as {pre-condition, action sequence, postcondition} triplets. By leveraging this structure we show that RL agents are able to succeed at the task.” |
“Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance”, by Bansal et al | SURPRISE Many analysts have claimed that the difficulty in explaining predictions and decisions made or recommended by deep learning algorithms makes humans less willing to trust and/or take action after receiving them. This new paper tests that claim. “Although the accuracy of Artificial Intelligence (AI) systems is rapidly improving, in many cases it remains risky for an AI to operate autonomously, e.g., in high-stakes domains or when legal and ethical matters prohibit full autonomy. A viable strategy for these scenarios is to form Human-AI teams, in which the AI system augments one or more humans by recommending its predictions, but the people retains agency and have accountability on the final decisions… “Many researchers have argued that such human-AI teams would be improved if the AI systems could explain their reasoning. In addition to increasing trust between human and machine, and between humans across the organization, one hopes that an explanation should help the responsible human know when to trust the AI’s suggestion and when to be skeptical, e.g., when the explanation doesn’t make sense. Such appropriate reliance is crucial for users to leverage AI assistance and improve task performance… “A careful analysis of existing literature reveals that prior studies observed improvements due to explanations only when the AI, alone, outperformed both the human and the best human-AI team. This raises an important question: can explanations lead to complementary performance, i.e., with accuracy higher than both the human and the AI working alone? “We address this question by devising comprehensive studies on human-AI teaming, where participants solve a task with help from an AI system without explanations and from one with varying types of AI explanation support. We carefully controlled to ensure comparable human and AI accuracy across experiments on three NLP datasets (two for sentiment analysis and one for question answering). “While we found complementary improvements from AI augmentation, they were not increased by state-of-the-art explanations compared to simpler strategies, such as displaying the AI’s confidence. We show that explanations increase the chance that humans will accept the AI’s recommendation regardless of whether the AI is correct.” |
“Why Tech Didn’t Save Us From COVID-19”, by David Rotman | “Although scientists identified and sequenced the new coronavirus within weeks of its appearance in late December—an essential step in creating a diagnostic—the US and other countries stumbled in developing PCR tests for general use. “Incompetence and a sclerotic bureaucracy at the US Centers for Disease Control meant the agency created a test that didn’t work and then insisted for weeks that it was the only one that could be used… "Combined with the lack of testing, a splintered and neglected system of collecting public health data meant epidemiologists and hospitals knew too little about the spread of the infection. “In an age of big data in which companies like Google and Amazon use all sorts of personal information for their advertising and shopping operations, health authorities were making decisions blind… “A once-healthy innovation ecosystem in the US, capable of identifying and creating technologies essential to the country’s welfare, has been eroding for decades… “The US has, over the last half-century, increasingly put its faith in free markets to create innovation. That approach has built a wealthy Silicon Valley and giant tech firms that are the envy of entrepreneurs around the world. But it has meant little investment and support for critical areas such as manufacturing and infrastructure—technologies relevant to the country’s most basic needs… “The problem with letting private investment alone drive innovation is that the money is skewed toward the most lucrative markets. The biggest practical uses of AI have been to optimize things like web search, ad targeting, speech and face recognition, and retail sales. “Pharmaceutical research has largely targeted the search for new blockbuster drugs. Vaccines and diagnostic testing, so desperately needed now, are less lucrative… |
| |
| |
| |
| |
| |
| |
| |
May20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Measuring the Algorithmic Efficiency of Neural Networks”, by Hernandez and Brown | SURPRISE This month, there are further indicators of the rate at which various technologies are developing. “Three factors drive the advance of AI: algorithmic innovation, data, and the amount of compute available for training. Algorithmic progress has traditionally been more difficult to quantify than compute and data… “We show that the number of floating-point operations required to [produce a standard benchmark result] has decreased by a factor of 44x between 2012 and 2019. This corresponds to algorithmic efficiency doubling every 16 months over a period of 7 years. Notably, this outpaces the original Moore’s law rate of improvement in hardware efficiency (11x over this period).” |
“The Changing Economics of Knowledge Production”, by Abis and Veldkamp | SURPRISE “Machine learning, artificial intelligence (AI), or big data all refer to new technologies that reduce the role of human judgment in producing usable knowledge… “Big data technologies change the way in which data and human labor combine to create knowledge. Is this a modest technological advance or a transformation of our basic economic processes? Using hiring and wage data from the financial sector, we estimate firms' data stocks and the shape of their knowledge production functions. “Knowing how much production functions have changed informs us about the likely long-run changes in output, in factor shares, and in the distribution of income, due to the new, big data technologies. Using data from the investment management industry, our results suggest that the labor share of income in knowledge work may fall from 44% to 27%”… “Our results inform us about how the demand for labor and data will change, how to value each in the new economy, and how the distribution of income is likely to shift, absent policy intervention.” |
“Explainable Artificial Intelligence: a Systematic Review”, by Vilone and Longo | Rapidly increasing use of machine learning (ML) and deep learning (DL) technologies in recent years has led to a matching increase in demand for technologies that can explain the logic behind their results (“Explainable Artificial Intelligence” or XAI). This new paper presents an extensive overview of the current state of the XAI field. The authors note that, “unfortunately, most of the models that have been built with ML and deep learning have been labeled ‘black-box’ by scholars because their underlying structures are complex, non-linear and extremely dicult to be interpreted and explained to laypeople. “This opacity has created the need for XAI architectures that is motivated mainly by three reasons: i) the demand to produce more transparent models; ii) the need of techniques that enable humans to interact with them; iii) the requirement of trustworthiness of their inferences. “Additionally, as proposed by many scholars, models induced from data must be liable as liability will likely soon become a legal requirement. Article 22 of the [European Union’s] General Data Protection Regulation (GDPR) sets out the rights and obligations of the use of automated decision making. Noticeably, it introduces the right of explanation by giving individuals the right to obtain an explanation of the inference/s automatically produced by a model, confront and challenge an associated recommendation, particularly when it might negatively affect an individual legally, financially, mentally or physically.” |
“From Probability to Consilience: How Explanatory Values Implement Bayesian Reasoning”, by Wojtowicz and DeDeo | SURPRISE This paper discusses in depth what we mean by “an explanation”. Along with “Explaining Explanation for Explainable AI” (which takes a naturalistic approach to explanation), and “Metrics for Explainable AI: Challenges and Prospects”, both by Hoffman, Klein, and Mueller, the current paper is the best one we have read on this increasingly critical topic. “Intuitively, philosophically, and as seen in laboratory experiments, explanations are judged as better or worse on the basis of many different criteria… “The multiplicity of values appears to conflict with Bayesian models of cognition, which speak solely in terms of degrees of beliefs and suggest we judge explanations as better or worse on the basis of a single quantity, the posterior likelihood [the extent to which the probability of observed data, given an explanation, aligns with observations].” The authors “show how to resolve these conflicts by arguing that previously-identified explanatory values capture different components of a full Bayesian calculation and, when considered together and weighed appropriately, implement Bayesian cognition.” These values include, “(i) descriptiveness, which measures the total extent to which the explanation predicts each fact in isolation from the others; (ii) co-explanation, which measures the extent to which the explanation links facts together; (iii) theoretical, or evidence-independent values; and (iv) context-dependent priors.” |
“Smarter enterprise search: why knowledge graphs and NLP can provide all the right answers”, by Accenture | SURPRISE This article outlines how improvements in two technologies are leading to substantial improvements in augmented cognition, and increasing the productivity of knowledge workers – but in the longer term, probably also leading to a reduction in their numbers. “The amount of information available to us is extraordinary. And it’s growing exponentially all the time: already amounting to 44 zetabytes, data volumes are predicted to hit 175 zetabytes in the next five years (IDC) . Eighty percent of this data is unstructured (emails, text documents, audio, video, social posts and so on), and just 20% is held in structured systems of some kind. “To find answers from this massive resource and pinpoint exactly what we’re looking for, we need a way to extract facts from documents and store those facts somewhere for easy access.” Rapid progress in two areas is making this much easier: Natural Language Processing (the automatic computational processing of human knowledge), and Knowledge Graphs. A Knowledge Graph is a model of a knowledge domain created by subject-matter experts with the help of intelligent machine learning algorithms (a process that is gradually being increasingly automated). It provides a structure and common interface for data and enables the creation of smart multilateral relations throughout your databases (based on a common subject-predicate-object format). The knowledge graph is essentially another layer that sits on top of an existing database, which in turn enables that use of artificial intelligence technologies to operate across multiple knowledge graphs, in applications such as advanced search. |
| |
| |
| |
| |
| |
| |
| |
| |
Apr20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Intelligent Automation: Getting Employees to Embrace the Bots”, by Heric et al from Bain & Company | COVID19 will very likely accelerate investment in automation. “Automation of business processes is rapidly scaling up, with fallout from the coronavirus likely accelerating adoption. Bain & Company’s survey of executives worldwide finds that companies report cost savings from automation of roughly 20% on average over the past two years. Other reported benefits include improved process quality and accuracy, reduced cycle times and improved compliance”… “But the path to benefits is bumpy. Some 44% of respondents said their automation projects have not delivered the expected savings. The major barriers all involve execution—notably, competing business priorities, insufficient resources or lack of skill”… “Fallout from the coronavirus outbreak may change this. As companies lose critical staff and the fragility of manual business processes is exposed, many companies will have no choice but to turn to automation to keep the business running. “Bain expects that by the end of the 2020s, automation of business processes may eliminate 20% to 25% of current jobs, hitting lower-skilled workers the hardest, and benefiting highly skilled workers and the owners of capital. In the US service sector, for instance, automation could spread through companies two to three times more rapidly than in previous transformations in agriculture, manufacturing and construction.” |
“A non-cooperative meta-modeling game for automated third-party calibrating, validating, and falsifying constitutive laws with parallelized adversarial attacks”, by Wang et al | Building models that enable the identification of solutions that are robust to a wide range of possible scenarios that can be produced by highly complex systems is a critical challenge for making good decisions in the face of uncertainty. The extent to which AI methods can be used to meet this challenge is a critical issue. While technical, this paper shows that significant progress is being made in this area. “As constitutive models that predict material responses become increasingly sophisticated and complex, the demands and difficulties for accurately calibrating and validating those constitutive laws also increase… “Engineering applications, particularly those involve high risk and high-regret decision-makings, require models to maintain robustness and accuracy in unforeseen scenarios using as little amount of necessary calibration data as possible…a common challenge is to detect rare events where a catastrophic loss of the prediction accuracy may occur in otherwise highly accurate constitutive laws. The authors “introduce concepts from game theory and machine learning techniques to overcome many of these existing difficulties…Competing AI agents systematically generate experimental data to calibrate a given constitutive model and to explore its weakness, in order to improve experiment design and model robustness through competition…By capturing all possible design options of the laboratory experiments into a single decision tree, we recast the design of experiments as a game of combinatorial moves that can be resolved through deep reinforcement learning by the two competing players.” |
“First return then explore”, by Ecoffet et al | This is an important development, that should enable reinforcement learning based models to more extensively explore their state space and avoid becoming trapped on “local optima”, or on a low peak when searching for the highest one on a fitness landscape. It is another example of how, out of range of the headline writers, AI technologies continue to improve in important ways. “Recent years have yielded impressive achievements in Reinforcement Learning (RL), including world-champion level performance in Go4, Starcraft II5, and Dota II6, as well as autonomous learning of robotic skills such as running, jumping, and grasping. Many of these successes were enabled by reward functions that are carefully designed to be highly informative. “However, for many practical problems, defining a good reward function is non-trivial… “A key observation is that sufficient exploration of the state space enables discovering sparse rewards as well as avoiding deceptive local optima. We argue that two major issues have hindered the ability of previous algorithms to explore: detachment, in which the algorithm loses track of interesting areas to explore from, and derailment, in which the exploratory mechanisms of the algorithm prevent it from returning to previously visited states, preventing exploration directly and/or forcing practitioners to make exploratory mechanisms so minimal that effective exploration does not occur. “We present Go-Explore, a family of algorithms designed to explicitly avoid detachment and derailment. We demonstrate how the Go-Explore paradigm allows the creation of algorithms that thoroughly explore environments.” |
“AutoML-Zero: Evolving Machine Learning Algorithms From Scratch”, by Real et al from Google | This is yet another indicator of the improving ability of AI technologies to automatically search through complex state spaces to identify robust solutions to specified problems – in this case, the development of new algorithms. “In recent years, neural networks have reached remarkable performance on key tasks and seen a fast increase in their popularity. This success was only possible due to decades of machine learning (ML) research into many aspects of the field, ranging from learning strategies to new architectures. The length and difficulty of ML research prompted a new field, named AutoML that aims to automate such progress by spending machine compute time instead of human research time. “This endeavor has been fruitful but, so far, modern studies have only employed constrained search spaces heavily reliant on human design…To address this, we propose to automatically search for whole ML algorithms using little restriction on form and only simple mathematical operations as building blocks. We call this approach AutoML-Zero, following the spirit of previous work that aims to learn with minimal human participation. “In other words, AutoML-Zero aims to search a fine-grained space simultaneously for the model, optimization procedure, initialization, and so on, permit- ting much less human-design and even allowing the discovery of non-neural network algorithms… “It is possible today to automatically discover complete machine learning algorithms just using basic mathematical operations as building blocks. We demonstrate this by introducing a novel framework that significantly reduces human bias through a generic search space. Despite the vastness of this space, evolutionary search can still discover two-layer neural networks trained by backpropagation. These simple neural networks can then be surpassed by evolving directly on tasks of interest.” |
“Flexible and Efficient Long Range Planning Through Curious Exploration”, by Curtis et al | Another indicator of AI progress: Curious exploration of state spaces to support more effective long range planning. “Many complex behaviors such as cleaning a kitchen, organizing a drawer, or cooking a meal require plans that are a combination of low-level geometric manipulation and high-level action sequencing. Boiling water requires sequencing high-level actions such as fetching a pot, pouring water into the pot, and turning on the stove. In turn, each of these high-level steps consists of many low-level task specific geometric action primitives. For instance, grabbing a pot requires intricate motor manipulation and physical considerations such as friction, force, etc. “The process of combining low-level geometric decisions and high-level action sequences is often referred to as multi-step planning. While high-level task planning and low-level geometric planning are difficult problems on their own, integrating them presents unique challenges that add further complexity… “Identifying algorithms that flexibly and efficiently discover temporally-extended multi-phase plans is an essential step for the advancement of robotics and model-based reinforcement learning. “The core problem of long-range planning is finding an efficient way to search through the tree of possible action sequences…we propose the Curious Sample Planner (CSP), which fuses elements of TAMP (Task and Motion Planning) and DRL (Deep Reinforcement Learning) by combining a curiosity-guided sampling strategy with imitation learning to accelerate planning. “We show that CSP can efficiently discover interesting and complex temporally-extended plans for solving a wide range of physically realistic 3D tasks. In contrast, standard planning and learning methods often fail to solve these tasks at all or do so only with a huge and highly variable number of training samples… “We [also] show that CSP supports task transfer so that the exploration policies learned during experience with one task can help improve efficiency on related tasks.” |
Education systems around the world are struggling with the switch to remote learning necessitated by the arrival of #COVID19 and the closure of schools. Going forward, the pandemic will very likely reduce birth rates and immigration. Faced with slower labor force growth, renewed economic growth (which will be critical to repaying increased debt) will therefore heavily depend on increased productivity growth. And that must largely come through a combination of increased use of AI and automation and higher quality labor inputs. However the latter will not happen without very substantial improvements in education systems that in many places have long resisted change despite stagnant or declining results (there are exceptions, like Alberta and Massachusetts). | Even before COVID19 arrived, a recent report by the World Bank concluded that, “in most countries, education systems are not providing workers with the skills necessary to compete in today’s job markets. The growing mismatch between demand and supply of skills holds back economic growth and undermines opportunity” (“The Learning Challenge in the 21st Century”, by Harry Patrinos). It remains to be seen whether COVID19 will cause that to change. In some places, unions are in effect demanding a return to the status quo ante (e.g., “Online School Demands More of Teachers. Unions Are Pushing Back” by Goldstein and Shapiro). At the same time, some (but not all) parents and students have found even quickly cobbled together forms of remote learning preferable to traditional “seat time” based schools. This has created a large potential market for new approaches to schooling that could at last break through the institutional constraints that have thus far held back the substantial improvements in education system performance.” |
It also seems clear that COVID19 will provide the long-needed impetus to improve many healthcare systems, whose shortcomings are now painfully apparent. These include poor management and lack of critical supplies; precarious finances as patient use of many services plummeted; and insufficient integration of acute, chronic, and social care systems that has resulted in very high death rates in the latter. | As in the case of education, necessary improvements in national healthcare systems will very likely be resisted by those with vested interests in the status quo. And yet they are equally critical, not only for improving effectiveness and efficiency of healthcare delivery in the resource constrained post-COVID19 age, but also to improve those systems’ resilience and ability to adapt to the next pandemic (e.g., H7N9 or H5N1 influenza). |
“A performance comparison of eight commercially available automatic classifiers for facial affect recognition”, by Dupre et al | This is an interesting indicator not only of the current state of video surveillance technologies, but also of the extent of their potential effectiveness in social control applications by authoritarian governments. “The ability to accurately detect what other people are feeling is an important element of social interaction. Only if we can perceive the affective state of an individual, will we be able to communicate in a way that corresponds to that experience. In the quest for finding a ‘window to the soul’ that reveals a view onto another’s emotion, the significance of the face has been a focus of popular and scientific interest alike… “In the wake of rapid advances in automatic affect analysis, commercial automatic classifiers for facial affect recognition have attracted considerable attention in recent years. While several options now exist to analyze dynamic video data, less is known about the relative performance of these classifiers, in particular when facial expressions are spontaneous rather than posed. In the present work, we tested eight out-of-the-box automatic classifiers, and compared their emotion recognition performance to that of human observers. “A total of 937 videos were sampled from two large databases that conveyed the basic six emotions (happiness, sadness, anger, fear, surprise, and disgust) either in posed or spontaneous form. “Results revealed a recognition advantage for human observers over automatic classification. Among the eight classifiers, there was considerable variance in recognition accuracy ranging from 48% to 62%. Subsequent analyses per type of expression revealed that performance by the two best performing classifiers approximated those of human observers, suggesting high agreement for posed expressions. However, classification accuracy was consistently lower (although above chance level) for spontaneous affective behavior. The findings indicate potential shortcomings of existing out-of-the-box classifiers for measuring emotions, and highlight the need for more spontaneous facial databases that can act as a benchmark in the training and testing of automatic emotion recognition systems.” |
| |
| |
| |
| |
| |
Mar20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“A New Infection Alarm System on Your Smartphone” by Rosenbach and Schmundt in Der Spiegel | COVID19 will undoubtedly accelerate the development and deployment of health related wearables and apps. These will likely reflect different privacy preferences and regulations around the world. This article is an example of a European approach, which is at one end of the privacy scale, while China’s approach is at the other. Expect US companies to fall somewhere in between. “In the debate over the deployment of digital technologies to help combat the spread of novel coronavirus, a new European approach has been developed in the hopes of diffusing fears of far-reaching surveillance via so-called ‘tracking apps.’ In recent weeks, a team of around 130 people from 17 institutes, organizations and companies in Europe has developed a technology that is intended as an alternative to the tracking technologies used in some countries in Asia. These technologies are sometimes applied against the will of users and some of them can enable other users to identify those who have become infected with the virus. “But the project PEPP-PT, which stands for Pan-European Privacy Protecting Proximity Tracing), seeks to avoid those pitfalls. The plan calls for users to voluntarily download an app which will inform them if they have recently been in the proximity of someone who subsequently tested positive for coronavirus and who also uses the app. That is the extent of the information that will be supplied: You were near someone who was later confirmed to be a carrier of the virus.” |
“China and Huawei propose reinvention of the internet”, by Gross and Murgia in the Financial Times | “China has suggested a radical change to the way the internet works to the UN, in a proposal that claims to enable cutting-edge technologies such as holograms and self-driving cars but which critics say will also bake authoritarianism into the architecture underpinning the web. “The telecoms group Huawei, together with state-run companies China Unicom and China Telecom, and the country’s Ministry of Industry and Information Technology (MIIT), jointly proposed a new standard for core network technology, called “New IP”, at the UN’s International Telecommunication Union (ITU). “The proposal has caused concerns among western countries including the UK, Sweden and the US, who believe the system would splinter the global internet and give state-run internet service providers granular control over citizens’ internet use.” |
“Multi-Agent Reinforcement Learning for Problems with Combined Individual and Team Reward”, by Seikh and Boloni, from the University of Central Florida | SURPRISE “Cooperative multi-agent problems are prevalent in real world settings such as strategic conflict resolution and coordination between autonomous vehicles… “Such problems can be modeled as dual-interest: each agent is simultaneously working towards maximizing its own payoff (local reward) as well as the collective success of the team (global reward)… “Learning multi-agent cooperation while simultaneously maximizing local rewards is still an open [technical] challenge” … The authors present “a novel cooperative multi-agent reinforcement learning framework that simultaneously learns to maximize the global and local rewards.” It substantially improves results compared to existing methods. |
“Learning Compositional Rules via Neural Program Synthesis”, by Nye et al | SURPRISE “Many aspects of human reasoning, including language, require learning rules from very little data. Humans can do this, often learning systematic rules from very few examples, and combining these rules to form compositional rule-based systems. “Current neural architectures, on the other hand, often fail to generalize in a compositional manner, especially when evaluated in ways that vary systematically from training. “In this work, we present a neuro-symbolic model which learns entire rule systems from a small set of examples. “Instead of directly predicting outputs from inputs, we train our model to induce the explicit system of rules governing a set of previously seen examples, drawing upon techniques from the neural program synthesis literature". |
“Is The Juice Worth The Squeeze? Machine Learning In and For Agent-Based Modeling”, by Dahlke et al | Along with development of causal and counterfactual reasoning capabilities, we believe that the integration of agent based modeling and deep learning has the potential to deliver very substantial improvements in the areas of understanding, explaining, predicting, and controlling the behavior of complex adaptive systems. Hence we carefully track surprises and indicators of progress in this area. In this paper, the authors “conduct a systematic literature review and classify the literature on the application of ML in and for ABM according to a theoretically derived classification scheme. We do so to investigate how exactly machine learning has been utilized in and for agent-based models so far and to critically discuss the combination of these two promising methods. “We find that, indeed, there is a broad range of possible applications of ML to support and complement ABMs in many different ways, already applied in many different disciplines. We see that, so far, ML is mainly used in ABM for two broad cases: First, the modelling of adaptive agents equipped with experience learning and, second, the analysis of outcomes produced by a given ABM.” |
“Deep Agent: Studying the Dynamics of Information Spread and Evolution in Social Networks”, by Garibay et al | SURPRISE In the US, the Defense Advanced Research Projects Agency (DARPA) “Computer Simulation of Online Social Behavior” (SocialSim) program aims to “ develop innovative technologies for high-fidelity computational simulation of online social behavior. SocialSim focuses specifically on information spread and evolution.” DARPA notes that, “Current computational approaches to social and behavioral simulation are limited in this regard. Top-down simulation approaches focus on the dynamics of a population as a whole, and model behavioral phenomena by assuming uniform or mostly-uniform behavior across that population. Such methods can easily scale to simulate massive populations, but can be inaccurate if there are specific, distinct variations in the characteristics of the population. “In contrast, bottom-up simulation approaches treat population dynamics as an emergent property of the activities and interactions taking place within a diverse population. While such approaches can enable more accurate simulation of information spread, they do not readily scale to represent large populations. SocialSim aims to develop novel approaches to address these challenges.” This paper “explains the design of a social network analysis framework, developed under DARPA’s SocialSim program, with novel architecture that models human emotional, cognitive and social factors… to uncover the underlying dynamics that explain the inner workings and reasons for the selection and diffusion of information in online social platforms.” It develops “a multi-resolution simulation at the user, community, population, and content levels.” |
“Knowledge Graphs”, by Hogan et al | This paper provides “a comprehensive introduction to knowledge graphs, which have recently garnered significant attention from both industry and academia in scenarios that require exploiting diverse, dynamic, large-scale collections of data” … A knowledge graph is “a graph of data intended to accumulate and convey knowledge of the real world, whose nodes represent entities of interest and whose edges represent relations between these entities.” Use of knowledge graphs “opens up a range of techniques than can be brought to bear for integrating and extracting value from diverse sources of data.” |
“A scalable pipeline for designing reconfigurable organisms”, by Kriegman et al | SURPRISE But for the arrival of COVID19, this would have been (and still should be) a much bigger story. For the first time, a team has produced “living robots” – that is, living, programmable organisms. Not only does this raise a host of ethical issues, but it also makes so-called “grey goo” risks posed by advancing nanotechnology – either from accident or intentional misuse – one step closer. “Most technologies are made from steel, concrete, chemicals, and plastics, which degrade over time and can produce harmful ecological and health side effects. It would thus be useful to build technologies using self-renewing and biocompatible materials, of which the ideal candidates are living systems themselves.” “Thus, we here present a method that designs completely biological machines from the ground up: computers automatically design new machines in simulation, and the best designs are then built by combining together different biological tissues.” “This suggests others may use this approach to design a variety of living machines to safely deliver drugs inside the human body, help with environmental remediation, or further broaden our understanding of the diverse forms and functions life may adopt.” |
COVID19 has highlighted many shortcomings in national healthcare and educational systems | In healthcare, the poor integration of the social care (nursing homes, assisted living, home health supports), particularly for the elderly, and the healthcare (preventative, chronic, and acute medical care) has become painfully apparent, with many COVID19 deaths in the social care system. The “traditional” healthcare system has also faced challenges. Yet here the contrast between different national healthcare systems is also painfully apparent, and will likely lead to reforms (yet in some countries, like the US, political conflicts that pre-date COVID19 may still block them, with very unpredictable medium term political effects). In education, the need to rapidly move to remote learning has highlighted the consequences of previous system design decisions. For example, systems that utilize a common curriculum have found it easier to switch to online learning. But even there, COVID19 has highlighted the need to change processes, staff skills, and structures. For example, a wholesale shift to remote learning has created an unprecedented opportunity to leverage the skills of most talented teachers, with other teachers focusing on supporting small groups and individual students. Yet that is running headlong into union contract provisions and K12 cultural norms. If COVID19 forces an extension of remote learning, it will raise even more difficult issues related to education staffing levels, facility needs, and budget reallocations. |
| |
| |
| |
| |
Feb20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence” by Gary Marcus | SURPRISE This is an outstanding paper that provides a great overview of key uncertainties in future improvements in AI capabilities. “Recent research in artificial intelligence and machine learning has largely emphasized general-purpose learning and ever-larger training sets and more and more [computational resources]. In contrast, I propose a hybrid, knowledge-driven, reasoning-based approach, centered around cognitive models, that could provide the substrate for a richer, robust AI than is currently possible… “Let us call that new level robust artificial intelligence: intelligence that, while not necessarily superhuman or self-improving, can be counted on to apply what it knows to a wide range of problems in a systematic and reliable way, synthesizing knowledge from a variety of sources such that it can reason flexibly and dynamically about the world, transferring what it learns in one context to another, in the way that we would expect of an ordinary adult. “In a certain sense, this is a modest goal, neither as ambitious or as unbounded as "superhuman" or "artificial general intelligence" but perhaps nonetheless an important, hopefully achievable, step along the way—and a vital one, if we are to create artificial intelligence we can trust, in our homes, on our roads, in our doctor's offices and hospitals, in our businesses, and in our communities. “Quite simply, if we cannot count on our AI to behave reliably, we should not trust it… “One might contrast robust AI with, for example, narrow intelligence, systems that perform a single narrow goal extremely well (e.g., chess playing or identifying dog breeds) but often in ways that are extremely centered around a single task and not robust and transferable to even modestly different circumstances (e.g., to a board of different size, or from one video game to another with the same logic but different characters and settings) without extensive retraining. “Such systems often work impressively well when applied to the exact environments on which they are trained, but we often can't count on them if the environment differs, sometimes even in small ways, from the environment on which they are trained. Such systems have been shown to be powerful in the context of games, but have not yet proven adequate in the dynamic, open-ended flux of the real world.” |
“How to Know if Artificial Intelligence is About to Destroy Civilization” by Oren Etzioni, CEO of the Allen Institute for AI | SURPRISE Etzioni proposes three “canaries in the coal mind” that would indicate accelerating AI capabilities. “First, In contrast to machine learning, human learning maps a personal motivation (“I want to drive to be independent of my parents”) to a strategic learning plan (“Take driver’s ed and practice on weekends”). A human formulates specific learning targets (“Get better at parallel parking”), collects and labels data (“The angle was wrong this time”), and incorporates external feedback and background knowledge (“The instructor explained how to use the side mirrors”). Humans identify, frame, and shape learning problems. “None of these human abilities is even remotely replicated by machines. Machines can perform superhuman statistical calculations, but that is merely the last mile of learning. The automatic formulation of learning problems, then, is our first canary. It does not appear to be anywhere close to dying… “Self-driving cars are a second canary. They are further in the future than anticipated by boosters like Elon Musk. AI can fail catastrophically in a typical situations, like when a person in a wheelchair is crossing the street. Driving is far more challenging than previous AI tasks because it requires making life-critical, real-time decisions based on both the unpredictable physical world and interaction with human drivers, pedestrians, and others… “AI doctors are a third canary. AI can already analyze medical images with superhuman accuracy, but that is only a narrow slice of a human doctor’s job. An AI doctor would have to interview patients, consider complications, consult other doctors, and more. These are challenging tasks that require understanding people, language, and medicine…it would have to approximate the abilities of human doctors across a wide range of tasks and unanticipated circumstances…Current AIs are idiots savants: successful on narrow tasks, such as playing Go or categorizing MRI images, but lacking the generality and versatility of humans.” |
“The New Business of AI (and How It’s Different from Traditional Software)”. Buy Casado and Bornstein, from Andreessen Horowitz | SURPRISE “At a technical level, artificial intelligence seems to be the future of software. AI is showing remarkable progress on a range of difficult computer science problems, and the job of software developers – who now work with data as much as source code – is changing fundamentally in the process. Many AI companies (and investors) are betting that this relationship will extend beyond just technology – that AI businesses will resemble traditional software companies as well. Based on our experience working with AI companies, we’re not so sure… “We have noticed in many cases that AI companies simply don’t have the same economic construction as software businesses. At times, they can even look more like traditional services companies. In particular, many AI companies have: 1. Lower gross margins due to heavy cloud infrastructure usage and ongoing human support; 2. Scaling challenges due to the thorny problem of edge cases; 3. Weaker defensive moats due to the commoditization of AI models and challenges with data network effects… The beauty of software (including Software as a Service, SaaS) is that it can be produced once and sold many times. This property creates a number of compelling business benefits, including recurring revenue streams, high (60-80%+) gross margins, and – in relatively rare cases when network effects or scale effects take hold – superlinear scaling. “Software companies also have the potential to build strong defensive moats because they own the intellectual property (typically the code) generated by their work. “Service businesses occupy the other end of the spectrum. Each new project requires dedicated headcount and can be sold exactly once. As a result, revenue tends to be non-recurring, gross margins are lower (30- 50%), and scaling is linear at best. Defensibility is more challenging – often based on brand or incumbent account control – because any IP not owned by the customer is unlikely to have broad applicability. “AI companies appear, increasingly, to combine elements of both software and services… Maintaining trained data models can feel, at times, more like a services business – requiring significant, customer-specific work and input costs beyond typical support and success functions.” |
“Technological interdependencies predict innovation dynamics”, by Pichler, Lafond, and Farmer | SURPRISE “Technological evolution is often described as a recursive process whereby the recombination of existing components leads to new or improved technological components…A simple hypothesis, therefore, is that technological domains that tend to recombine elements from fast-growing technological domains should themselves grow faster. In other words, a technology will tend to progress faster if the technologies it relies on are themselves making fast progress. While these ideas are well established, very little has been done to establish empirically that technological interdependencies help predict future innovation dynamics. “Being able to demonstrate this relationship would be very helpful, as it would allow us to support key technologies and overall technological progress by designing and supporting technological ecosystems… “We propose a simple model where the innovation rate of a technological domain depends on the innovation rate of the technological domains it relies on. Using data on US patents from 1836 to 2017, we make out-of-sample predictions and find that the predictability of innovation rates can be boosted substantially when network effects are taken into account. “In the case where a technology’s neighborhood future innovation rates are known, the average predictability gain is 28% compared to simpler time series model which do not incorporate network effects. Even when nothing is known about the future, we find positive average predictability gains of 20%. The results have important policy implications, suggesting that the effective support of a given technology must take into account the technological ecosystem surrounding the targeted technology.” |
“The US Intelligence Community is Caught in a Collector’s Trap”, by Zachery Brown, DefenseOne, 25Feb20 | This article is very much in line with what we've been saying at The Index Investor since 1997… “The information haystack in which we search for useful needles is growing faster than we could ever catch up. Gathering more hay isn't the answer… while the datasphere grows geometrically, the mechanisms intelligence services use to make sense of it—spies, listening posts, and satellites—can only be added arithmetically. The gap between information to collect and information that is actually collected keeps growing larger, and can never be closed…. But not for lack of trying. For decades, the U.S. intelligence community has added to its expansive data-collection enterprise. Today, it costs the taxpayer around $80 billion a year. National Intelligence University researcher Josh Kerbel calls this the community’s “classified collection business model.” It is premised upon the idea that information and intelligence are essentially synonymous, and posits that the time and treasure spent gathering and sorting it all is justified because it leads to better policy choices—what the community calls “decision advantage.” “The record of American foreign policy failures in my lifetime alone, however, suggests that this model is flawed…. “Even if it were possible to gather every bit of information relevant to national security (it isn’t), it wouldn’t serve policymakers as well as you might think. We tend to find the things we look for and are surprised when the things we are not looking for find us instead… "What makes the collector’s trap so insidious is not only its elusive goal but also the fact that the more information we do gather, the more confused we become. Human susceptibility to cognitive errors such as availability bias and the observer-expectancy effect means that with a virtually limitless amount of information already at our fingertips, certainty about practically anything has only decreased… "The intelligence community has over-invested in technical collection platforms at the expense of the people who give the information those systems collect context. Today’s consumers of intelligence are drowning in data, but thirsting for insight… "The intelligence community must chart a bold new model suited to the information-rich reality of our digital era, and finally, break free from the collector’s trap… "But because we can’t possibly collect everything, or even everything we think may be relevant, we must put far more emphasis on cultivating anticipation and foresight. We must become comfortable with uncertainty rather than trying to eliminate it. We must expect surprise, and grow more resilient, adaptive organizational structures and policies better suited to endure and incorporate the lessons learned from them. "At base, intelligence leaders must remind themselves that they are not in the business of collecting and protecting information, but of delivering insight and facilitating understanding so that better decisions can be made to advance national interests." |
| |
| |
| |
| |
| |
| |
| |
| |
Jan20: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Researchers: Are we on the cusp of an AI winter?”, by Sam Shead in Technology Review | “Hype surrounding AI has peaked and troughed over the years as the abilities of the technology get overestimated and then re-evaluated… “The peaks are known as AI summers, and the troughs AI winters. The 2010s were arguably the hottest AI summer on record with tech giants repeatedly touting AI’s abilities…There are signs, however, that the hype might be about to start cooling off…At the end of 2019, the smartest computers could still only excel at a "narrow" selection of tasks… “Gary Marcus, an AI researcher at New York University, said: ‘By the end of the decade there was a growing realisation that current techniques can only carry us so far.’ He thinks the industry needs some "real innovation" to go further… “Another researcher noted that, ‘One of the biggest challenges is to develop methods that are much more efficient in terms of the data and compute power required to learn to solve a problem well. In the past decade, we've seen impressive advances made by increasing the scale of data and computation available, but that's not appropriate or scalable for every problem. ‘If we want to scale to more complex behaviour, we need to do better with less data, and we need to generalise more.’" |
“The Unavoidable” by Eric Hanushek from Stanford (in our experience, one of the most astute observers of the true state of K12 education in the United States) | Few people appreciate the size of the negative second and third order effects that will result from the cumulative impact of an education system that is improving slowly, if at all, while labor replacing automation and artificial intelligence technologies continue to exponentially improve. As Hanushek writes, “education strongly affects the future economic returns that individuals see. It also dictates where the US economy will go in the future. Unfortunately, students in the United States are not competitive with students from much of the developed world… “The results of the 2019 National Assessment of Educational Progress (NAEP) underscore the serious (and frustrating) achievement problems facing the United States. They represent real problems that affect not only the children of this generation but also the future economies of all states. “These are not problems that can be put off. The burden on the United States will increase over time, any solutions will necessarily take time, and delay will exacerbate the problems…Not dealing effectively with these problems will cause increasing economic displacement as new technologies continue to replace workers with automation”. |
“The Impact of Artificial Intelligence on the Labor Market”, by Michael Webb | SUPRRISE “Artificial intelligence, or machine learning, refers to algorithms that learn to complete tasks by identifying statistical patterns in data, rather than following instructions provided by humans…At a time when rising inequality is a major social and political issue, it is unclear whether AI will increase inequality by, say, further displacing production workers, or reduce it by displacing doctors and lawyers.” The author “develops a new method for identifying which tasks can be automated by any particular technology…based on the following key idea. The text of patents contains information about what technologies do, and the text of job descriptions contains information about the tasks people do in their jobs. "These two text corpuses can be combined to quantify how much patenting in a particular technology has been directed at the tasks of any particular occupation. This is therefore a measure of the tasks from which labor may be displaced”… “Patents describe artificial intelligence performing tasks such as predicting prognosis and treatment, detecting cancer, identifying damage, and detecting fraud. These are tasks involved in medical imaging and treatment, insurance adjusting, and fraud analysis, all areas that are currently seeing high levels of AI research and development. “Notice that these activities are of a very different kind to those identified for robots and software. Whereas robots perform “muscle” tasks and software performs routine information processing, AI performs tasks that involve detecting patterns, making judgments, and optimization… Webb finds that, “high-skill occupations are most exposed to AI, with exposure peaking at about the ninetieth percentile. While individuals with low levels of education are somewhat exposed to AI, it is those with college degrees, including Master’s degrees, who are most exposed. Moreover, as might be expected from the fact that AI-exposed jobs are predominantly those involving high levels of education and accumulated experience, it is older workers who are most exposed to AI, with younger workers much less so… “These descriptive results clearly indicate that AI will affect very different occupations, and so different kinds of people, than software and robots.” |
“Extending the Race Between Education and Technology”, by Autor et al | Two phenomena are affecting relative wages in the United States. The first is the “race between education and technology” – can the US education system provide enough skilled graduates (and reskilled workers) to either stave off automation and/or to fill the new jobs exponentially improving technology will create, which have significantly different skill requirements? The authors note that, “the race between education and technology [RBET] provides a canonical framework that does an excellent job of explaining US wage structure changes across the twentieth century. The framework involves secular increases in the demand for more-educated workers from skill-biased technological change, combined with variations in the supply of skills from changes in educational access… “Increased educational wage differentials explain 75 percent of the rise of U.S. wage inequality from 1980 to 2000 as compared to 38 percent for 2000 to 2017… “A great economic divide has emerged between college-educated workers and those with less than a college degree. Ever since 1980, educational wage differentials have greatly expanded, and soaring income inequality has deeply marked the US economy… “Educational wage gains and overall wage and income inequality have closely followed changes in educational attainment against a backdrop of increased relative demand for more-educated workers from skill-biased technological change (SBTC). The implicit framework is one of a race between education and technology (RBET)… The idea is that there is secular growth in the demand for more-educated workers from SBTC at the same time there is rapid, but variable, growth of the relative supply of more-educated workers [and the relative capabilities of potentially labor displacing technology]… From 1980 to 2005, a slowdown in relative education supply growth contributed to a soaring college wage premium. That’s the saga of educational wage differentials from the 1890s to 2005… [However], “most of the recent rise in wage inequality has occurred within, rather than between, education groups. In fact, the largest part of increased wage variance in the twenty-first century comes from rising inequality among college graduates. There is almost no change in wage inequality for non-college workers since 2000. Such a pattern is consistent with the continuing, rapid rise of the 90-50 [percentile] wage differential and soaring top end inequality, combined with stability in the 50-10 wage differential in the 2000s. “Comprehending rising wage inequality in the 2000s requires a better understanding of growing wage inequality among college graduates, the rise in the return to post-BA education, and stagnant earnings of middle-wage workers (the upper half of non-college and lower half of college workers). The RBET framework remains relevant in the twenty-first century, but needs some tweaks.” |
“Artificial Intelligence and the Manufacturing of Reality” by Paul and Posard from RAND | “In 2016, a third of surveyed Americans told researchers they believed the government was concealing what they knew about the “North Dakota Crash,” a conspiracy made up for the purposes of the survey by the researchers themselves. This crash never happened, but it highlights the flaws humans carry with them in deciding what is or is not real. “The internet and other technologies have made it easier to weaponize and exploit these flaws, beguiling more people faster and more compellingly than ever before. It is likely artificial intelligence will be used to exploit the weaknesses inherent in human nature at a scale, speed, and level of effectiveness previously unseen. “Adversaries like Russia could pursue goals for using these manipulations to subtly reshape how targets view the world around them, effectively manufacturing their reality. If even some of our predictions are accurate, all governance reliant on public opinion, mass perception, or citizen participation is at risk… “One characteristic human foible is how easily we can falsely redefine what we experience. This flaw, called the ‘Thomas Theorem’, suggests, ‘If men define situations as real, they are real in their consequences.’ Put another way, humans not only respond to the objective features of their situations but also to their own subjective interpretations of those situations, even when these beliefs are factually wrong. “Other shortcomings include our willingness to believe information that is not true and a propensity to be as easily influenced by emotional appeals as reason, as demonstrated by the ‘North Dakota Crash’ falsehood. “Machines can also be taught to exploit these flaws more effectively than humans: Artificial intelligence algorithms can test what content works and what does not over and over again on millions of people at high speed, until their targets react as desired… “Defending against such massive manipulation will be particularly tricky given the current social media landscape, which allows for the easy multiplication of inauthentic individuals, personas, and accounts through the use of bots or other forms of automation.” |
| |
| |
| |
| |
| |
| |
| |
| |
Dec19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
Hearing of the US House of Representatives Committee on Financial Services Task Force on Artificial Intelligence on the Impact of AI on Capital Markets and Jobs in the Financial Services Industry | In his testimony, Dr. Marcos Lopez de Prado from Cornell University began by noting that, “as a consequence of recent advances in pattern recognition, big data and supercomputing, ML can today accomplish tasks that until recently only expert humans could perform… “ML algorithms are particularly powerful at modeling complex non-linear interactions between variables…According to studies, more that 34% of the total hedge fund assets under management are currently invested using algorithmic strategies, for over $1 trillion dollars. This figure does not include factor-based mutual funds and exchange traded funds offered to retail investors, so the total assets of algorithmic-managed investments could be close to $2 trillion…Eventually, we can expect that ML algorithms will be involved in the allocation of tens of trillions of dollars, replacing human discretion and the more traditional econometric methods… “Financial ML creates a number of challenges for the 6.14 million people employed in the finance and insurance industry, many of whom will lose their jobs, not necessarily because they are replaced by machines, but because they are not trained to work alongside algorithms. The retraining of these workers is an urgent and difficult task.” In separate testimony, Rebecca Fender from the CFA Institute noted that, “The type of skill required for investment teams will remain predominantly investment skills. Professionals on investment teams who understand the basics of AI, data science, and technology, however, can be expected to be far more effective than someone with similar investment skills but no exposure to such technologies…Today just 6% of CFA members and candidates say they are proficient in data analysis coding (Python, R, MATLAB, etc.)” |
“How connected is too connected? Impact of network topology on systemic risk and collapse of complex economic systems” by Vie and Morales | SURPRISE This paper highlights how, across many socio-technical systems, the increase in network connectivity over the past decade has likely made them much more exposed to systemic failure risks. The authors note that, “economic interdependencies have become increasingly present in globalized pro-duction, financial and trade systems. While establishing interdependencies among economic agents is crucial for the production of complex products, they may also in- crease systemic risks due to failure propagation. It is crucial to identify how network connectivity impacts both the emergent production and risk of collapse of economic systems.” They “propose a model to study the effects of network structure on the behavior of economic systems by varying the density and centralization of connections among agents. The complexity of production increases with connectivity given the combinatorial explosion of parts and products.” The authors find that, “emergent systemic risks arise when interconnections increase vulnerabilities.” They claim that their results “suggest a universal description of economic collapse given in the emergence of tipping points and phase transitions in the relationship between network structure and risk of individual failure. This relationship seems to follow a [S-curve] in the case of increasingly denser or centralized networks. The model sheds new light on the trade-off between increasing the density of connections in, and potential production of, a system and its robustness to collapse…Sparser networks may have a lower productivity but are more resilient to probability of failure…Economic agents may in some situations become too interconnected to thrive.” |
“Evolving trading strategies in heterogeneous environments”, by Dewhurst et al | SURPRISE “Securities markets are quintessential complex adaptive systems in which heterogeneous agents compete in an attempt to maximize returns. Species of trading agents are also subject to evolutionary pressure as entire classes of strategies become obsolete and new classes emerge. Using an agent-based model of interacting heterogeneous agents as a flexible environment that can endogenously model many diverse market conditions, we subject deep neural networks to evolutionary pressure to create dominant trading agents…and construct a method to turn [them] into trading algorithms. “We backtest these trading algorithms on real high-frequency foreign exchange data, demonstrating that elite trading algorithms are consistently profitable in a variety of market conditions, even though these algorithms had never before been exposed to real financial data. “These results provide evidence to suggest that developing trading strategies by repeated simulation and evolution in a mechanistic market model may be a practical alternative to explicitly training models with past observed market data.” |
The Artificial Intelligence Index, 2019 Annual Report | “In a year and a half, the time required to train a large image classification system on cloud infrastructure has fallen from about three hours in October 2017 to about 88 seconds in July, 2019. During the same period, the cost to train such a system has fallen similarly… “Progress on some broad sets of natural-language processing classification tasks…has been remarkably rapid; [however] performance is still lower on some NLP tasks requiring reasoning, or human-level concept learning tasks… “Prior to 2012, AI results closely tracked Moore’s Law, with compute doubling every two years. Post-2012, compute has been doubling every 3.4 months… “Globally, investment in AI startups continues its steady ascent. From a total of $1.3B raised in 2010 to over $40.4B in 2018 (with $37.4B in 2019 as of November 4th), funding has increased at an average annual growth rate of over 48%. |
“Data-driven Discovery of Emergent Behaviors in Collective Dynamics”, by Maggioni et al | SURPRISE “Particle and agent-based systems are a ubiquitous modeling tool in many disciplines. We consider the fundamental problem of inferring interaction [relationships] from observations of agent-based dynamical systems given observations of trajectories, in particular for collective dynamical systems exhibiting emergent behaviors with complicated interaction [relationships]… “We provide extensive numerical evidence that [our method for estimating these relationships] provides faithful approximations [of them], and provides accurate predictions for trajectories started at new initial conditions, both throughout the training" time interval in which the observations were made, and often much beyond.” |
| |
| |
| |
| |
| |
| |
| |
| |
Nov19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Teacher Pay Raises Aren’t Enough” by Eric Hanushek from Stanford University | he OECD has reported the latest PISA scores. This international assessment measures 15 year olds’ ability to apply their knowledge of reading, math, and science. The US ranked 8th in reading, 11th in science, and 30th in math, despite higher spending on education. As a Financial Times headline noted, “Rich-nation education splurge fails to deliver results.” Unfortunately, improved education results are critical for raising labor productivity and offsetting the economic impact of ageing populations on the economic growth of many developed nations. Hanushek notes that, “The discussion educators and policymakers need to have begins With a simple fact: U.S. teachers are woefully underpaid…the penalty shows up in the quality of our teacher force [compared to other nations’]… Since today's students become our future labor force, their lower achievement bodes ill for the future economic well-being of our nation…if the United States is to improve its schools, the available research indicates that the only feasible solution is to increase the overall effectiveness of our teachers”… However, “higher salaries may work to retain the most effective teachers, but they also retain the least effective. In fact, the across-the-board raises—the hallmark of the settlements with teachers unions—make getting out of the bad salary situation even more difficult, because they slow the chances of any new openings in the teacher force”… “Additionally, raises going indiscriminately to ineffective teachers likely dampen public enthusiasm for salary increases.” Hanushek concludes that, “the only practical solution apparent to me is the "grand bargain"—an idea broached more than 15 years ago but now perhaps more feasible as teacher salaries stagnate and U.S. student achievement continues to lag that of other countries. This bargain is simple: a substantial increase in teacher salaries combined with policies that produce a significant tilt toward more effective teachers” [by making it easier to dismiss weak ones]. |
“Why the U.S. Innovation Ecosystem Is Slowing Down” by Arora et al | SURPRISE Many observers have long noted the declining interest of practitioners in the social science and business research being produced by universities, pointing with envy to the closer relationship they believe exists between practitioners and university researchers in science and engineering. This new paper finds that now a gap is growing there too. “Data from the National Science Foundation (NSF) indicate that U.S. investment in science has steadily increased between 1970 and 2010, as measured by dollars spent (which has gone up 5X), number of PhDs trained (2X) and articles published (7X). Why is there little productivity growth to show for this?” … “One explanation, which we explore, is that today’s science is not being translated into applications — in other words, something is keeping scientific discoveries from fueling productive innovation. Our research finds that the U.S. innovation ecosystem has splintered since the 1970s, with corporate and academic science pulling apart and making application of basic scientific discoveries more difficult… “This marks a significant shift... We’ve moved from an economy where big firms did both scientific research and development toward one with a starker division of labor, where corporations specialize in development, and universities specialize in research…University researchers are rewarded for precedence (“who comes first”), while corporate researchers are rewarded for their usefulness in invention (“does it work”). Therefore, university research is more likely to be new, but less likely to function as intended by businesses.” |
“Detecting And Quantifying Causal Associations In Large Nonlinear Time Series Datasets”, by Runge et al | SURPRISE Better integration of causality has been a major challenge for artificial intelligence developers. This research suggests that progress in this area may be accelerating. “Identifying causal relationships and quantifying their strength from observational time series data are key problems in disciplines dealing with complex dynamical systems such as the Earth system or the human body… "The goal in time series causal discovery from complex dynamical systems is to statistically reliably estimate causal links, including their time lags… "Data-driven causal inference in such systems is challenging since datasets are often high dimensional and nonlinear with limited sample sizes…We introduce a novel method that flexibly combines linear or nonlinear conditional independence tests with a causal discovery algorithm to estimate causal networks from large-scale time series datasets… “Our method outperforms state-of-the-art techniques in detection power, which opens up entirely new possibilities to discover and quantify causal networks from time series across a range of research fields.” |
“Global AI Survey: AI proves its worth, but few scale impact”, by McKinsey | The pace at which organizations implement artificial intelligence technologies is just as important as the rate at which those technologies improve. More specifically, the extent to which AI adoption results in higher unemployment, and the extent to which is widens the gap between highly profitable firms and others that are barely earning their cost of capital has major implications for the evolution of productivity, inequality, and social and political conflict. This latest McKinsey analysis notes that, “most companies report measurable benefits from AI where it has been deployed; however, much work remains to scale impact, manage risks, and retrain the workforce. A group of high performers shows the way… “Adoption of artificial intelligence (AI) continues to increase, and the technology is generating returns. The findings of the latest McKinsey Global Survey on the subject show a nearly 25 percent year-over-year increase in the use of AI in standard business processes, with a sizable jump from the past year in companies using AI across multiple areas of their business. A majority of executives whose companies have adopted AI report that it has provided an uptick in revenue in the business areas where it is used, and 44 percent say AI has reduced costs. “The results also show that a small share of companies—from a variety of sectors—are attaining outsize business results from AI, potentially widening the gap between AI power users and adoption laggards…our results suggest that workforce retraining will need to ramp up. While the findings indicate that AI adoption has generally had modest overall effects on organizations’ workforce size in the past year, about one-third of respondents say they expect AI adoption to lead to a decrease in their workforce in the next three years, compared with one-fifth who expect an increase.” |
“The Measure of Intelligence”, by Francois Chollet from Google | SURPRISE This paper will very likely help to clarify the ongoing debate about the rate of technological progress towards AI that represents “artificial general intelligence.” Chollet notes that, “to make deliberate progress towards more intelligent and more human-like artificial systems, we need to be following an appropriate feedback signal: we need to be able to define and evaluate intelligence in a way that enables comparisons between two systems, as well as comparisons with humans. “Over the past hundred years, there has been an abundance of attempts to define and measure intelligence, across both the fields of psychology and AI. We summarize and critically assess these definitions and evaluation approaches, while making apparent the two historical conceptions of intelligence that have implicitly guided them. “We note that in practice, the contemporary AI community still gravitates towards benchmarking intelligence by comparing the skill exhibited by AIs and humans at specific tasks, such as board games and video games. “We argue that solely measuring skill at any given task falls short of measuring intelligence, because skill is heavily modulated by prior knowledge and experience: unlimited priors or unlimited training data allow experimenters to “buy” arbitrary levels of skills for a system, in a way that masks the system’s own generalization power. “We then articulate a new formal definition of intelligence based on Algorithmic Information Theory, describing intelligence as skill-acquisition efficiency and highlighting the concepts of scope, generalization difficulty, priors, and experience, as critical pieces to be accounted for in characterizing intelligent systems. Using this definition, we propose a set of guidelines for what a general AI benchmark should look like.” |
| |
| |
| |
| |
| |
| |
| |
| |
Oct19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Machines Beat Humans in a Reading Test. But Do They Understand?” by Jon Fox for Quanta Magazine | Natural Language Processing (NLP) is a critical area of AI development. Software is now beginning to best humans on reading comprehension benchmark tests. But Fox explores a critical question: “is AI actually starting to understand our language, or is it just getting better at gaming our systems?” He concludes that for now, at least, it is likely the latter, noting that, no training data set, no matter how comprehensively designed or carefully filtered, can capture all the edge cases and unforeseen inputs that humans effortlessly cope with when we use natural language.” |
“The Current State of AI and Deep Learning”, by Gary Marcus | Marcus wrote this article on Medium as a reply to Yoshua Bengio, with whom he has been having a debate on the limitations of deep learning systems. Marcus notes the areas where they agree, including the difficulty of generalizing beyond training data sets, especially in evolving systems; the need to bring causality (per Judea Pearl) into artificial reasoning, the need to incorporate sequential or deliberative reasoning (Kahneman’s System II), and the need to incorporate prior knowledge. Marcus believes that progress in AI (beyond perceptual classification and prediction) heavily depends on the development of better systems for symbol manipulation (e.g., performing operations over variables, like relating unstructured text to existing knowledge, logical argument, or algebra). As he notes, “Mapping a set of entities onto a set of predetermined categories (as deep learning does well) is not the same as generative novel interpretations or formulating a plan that crosses multiple time scales. There is no particular reason to think that the deep learning can do the latter two sorts of problems well.” |
Results of the 2019 US National Assessment of Educational Progress were announced, and showed that performance has now remained stagnant for a decade. | Surprise Writing in National Review (“School Reform Struggles”), Jay Greene and Rick Hess noted that, “The U.S. is distinctive for its sprawling, decentralized system of schools, which are governed in large part by 50 legislatures and more than 14,000 democratically controlled school districts.” “This means that, for better or worse, educational improvement is always a political project. The failure to improve schooling is thus, in part, inevitably a political failure. After all, improving schools nationwide requires enacting reforms across an array of contexts, and then executing, supporting, and sustaining those reforms in a patchwork of red and blue communities. This Tocquevillian challenge can be answered only with a broad, bipartisan coalition. We suspect that the dismal results recorded by the NAEP are partially due to a once-bipartisan school-reform community’s hard turn to the left.” Unfortunately, continued poor K12 education performance has long term implications. The lack of talent will only accentuate the gulf between the performance of companies that have it and those that don’t, as talent is critical to absorbing and deploying rapidly improving automation and AI technologies. Similarly, a lack of talent will tempt more companies to pursue investment in those technologies to build “labor lite” business models. That will contribute to worsening inequality, and social and political conflict. |
“Science and Technology Advance Through Surprise”, by Shi and Evans | Surprise This new research finds that, “breakthrough discoveries and inventions involve unexpected combinations of contents including problems, methods, and natural entities, and also diverse contexts such as journals, subfields, and conferences. Drawing on data from tens of millions of research papers, patents, and researchers, we construct models that predict more than 95% of next year’s content and context combinations…[We find that] breakthroughs occur when problems in one field are unexpectedly solved by researcher from another”… “We show how surprising successes systematically emerge across, rather than within communities of researchers; most commonly when those in one field surprisingly publish problem-solving results to audiences in a distant other.” |
“Long-range Event-level Prediction and Response Simulation for Urban Crime and Global Terrorism with Granger Networks”, by Li et al | Surprise “Large-scale trends in urban crime and global terrorism are well-predicted by socio-economic drivers, but focused, event-level predictions have had limited success” … Standard machine learning approaches are promising, but lack interpretability, are generally interpolative, and ineffective for precise future interventions with costly and wasteful false positives. Such attempts have neither adequately connected with social theory, nor analyzed disparities between urban crime and differentially motivated acts of societal violence such as terrorism. Thus, robust event level predictability is still suspect, and policy optimization via simulated interventions remains unexplored.” The authors “introduce Granger Network inference as a new forecasting approach for individual infractions with demonstrated performance far surpassing past results, yet transparent enough to validate and extend social theory.” |
“What’s Behind the Technological Hype” by Jeffrey Funk | “The percentage of start-up companies in the United States that are profitable at the time of their initial public stock offering has dropped to levels not seen since the 1990s dotcom stock market bubble…the large losses are easily explained: extreme levels of hype about new technologies, and too many investors willing to believe it…” Funk “discuss economic data showing that many highly touted new technologies are seriously over-hyped, a phenomenon driven by online news and the professional incentives of those involved in promoting innovation and entrepreneurship. This hype comes at a cost—not only in the form of record losses by start-ups, but in their inability to pursue alternative designs and find more productive and profitable opportunities, and in the inability of America’s decision-makers to acknowledge that innovation has slowed.” |
| |
| |
| |
| |
| |
| |
| |
Sep19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
The 12Sep19 edition of The Economist had a special section on the next revolution in technology. | Some highlights: “Drastic falls in cost are powering another computer revolution…The Internet of Things is the next big idea in computing.” “Over the past century electricity has allowed consumers and businesses at least in the rich world, access to a fundamental, universally useful good—energy—when and where they needed it. The IoT aims to do for information what electricity did for energy.” “The computerisation of everything is a big topic, and one that will take decades to play out. The result will be a slow-burning revolution of quantifiability in which knowledge that used to be fuzzy or incomplete or even non-existent becomes increasingly precise. That will give rise to what sports coaches call “marginal gains”. A 10% decrease in costs or a 15% cut in energy use are individually unexciting. Put enough of them together, though, and they will amount to a revolution in productivity…” |
“Augmented Data Science: Towards Industrialization and Democratization of Data Science”, by Uzunalioglu et al from Nokia Bell Labs | Surprise If deployed at scale, this technique could substantially speed the deployment and implementation of data science methods. “Conversion of raw data into insights and knowledge requires substantial amounts of effort from data scientists. Despite breathtaking advances in Machine Learning (ML) and Artificial Intelligence (AI), data scientists still spend a majority of their effort in understanding and then preparing the raw data for ML/AI. The effort is often manual and ad hoc, and requires some level of domain knowledge. The complexity of the effort increases dramatically when data diversity, both in form and context, increases. “We introduce our solution, Augmented Data Science (ADS), towards addressing this “human bottleneck” in creating value from diverse datasets. ADS is a data-driven approach and relies on statistics and ML to extract insights from any data set in a domain-agnostic way to facilitate the data science process.” |
“Google Claims to Have Reached Quantum Supremacy”, Financial Times, 20Sep19 | Surprise “Google claims to have built the first quantum computer that can carry out calculations beyond the ability of today’s most powerful supercomputers, a landmark moment that has been hotly anticipated by researchers. “A paper by Google’s researchers seen by the FT, that was briefly posted earlier this week on a Nasa website before being removed, claimed that their processor was able to perform a calculation in three minutes and 20 seconds that would take today’s most advanced classical computer, known as Summit, approximately 10,000 years…” “The researchers said this meant the “quantum supremacy”, when quantum computers carry out calculations that had previously been impossible, had been achieved…. “The Google researchers called it ‘a milestone towards full-scale quantum computing’. They also predicted that the power of quantum machines would expand at a “double exponential rate”, compared to the exponential rate of Moore’s Law, which has driven advances in silicon chips in the first era of computing.” See also, “The Next Decade in Quantum Computing – and How to Play”, by Gerbert and Ruess from BCG, and, predictably, “Rivals rubbish Google’s claim of quantum supremacy” by Richard Waters (FT 23Sep19) |
“Superhuman AI for Multiplayer Poker” by Brown and Sandholm in Science | Surprise The difficulty of the challenge met by Pluribus shows how AI technology can now be applied to situations involving complex strategic interactions between multiple parties. “The past two decades have witnessed rapid progress in the ability of AI systems to play increasingly complex forms of poker. However, all prior breakthroughs have been limited to settings involving only two players. Developing a superhuman AI for multiplayer poker was the widely-recognized main remaining milestone.” This paper describes, “Pluribus, an AI capable of defeating elite human professionals in six-player no-limit Texas hold’em poker, the most commonly played poker format in the world...” “Multiplayer games present fundamental additional issues beyond those in two-player games, and multiplayer poker is a recognized AI milestone… “The core of Pluribus’s strategy was computed via self play [i.e., the use of generative adversarial networks, or GANS], in which the AI plays against copies of itself, without any data of human or prior AI play used as input. The AI starts from scratch by playing randomly, and gradually improves as it determines which actions, and which probability distribution over those actions, lead to better outcomes against earlier versions of its strategy…” “As Jonathan Russo noted in The Observer, Texas hold’em is the epitome of multi-player strategic thinking, cue analysis, advanced future predictability and the most powerful, yet unquantifiable, skill of all, bluffing…The shocking fact that Pluribus out-bluffed the world’s top five players and won a big pot of money, means AI has reached a new critical level, because bluffing is one of the traits that enables us to succeed or fail by processing, and then utilizing, what we have learned in the past about the behaviors of others and thus affect the future to our advantage…. “The race to build better bluffing programs will be brutal. If a machine can out-bluff human traders or analysts, the foundation of the financial industry will be undermined.” |
“Are You Developing Skills that Won’t Be Automated?” by Stephen Kosslyn in HBR 25Sep19 | This is a thoughtful article that addresses a critical question more and more people are asking themselves. “After reflecting on characteristics of numerous jobs and professions, two non-routine kinds of work seem to me to be particularly common, and difficult to automate: “First, emotion. Emotion plays an important role in human communication…It is critically involved in virtually all forms of nonverbal communication and in empathy. But more than that, it is also plays a role in helping us to prioritize what we do, for example helping us decide what needs to be attended to right now as opposed to later in the evening… “Emotion is not only complex and nuanced, it also interacts with many of our decision processes. The functioning of emotion has proven challenging to understand scientifically (although there has been progress), and is difficult to build into an automated system...” “Second, context. Humans can easily take context into account when making decisions or having interactions with others. Context is particularly interesting because it is open ended— for instance, every time there’s a news story, it changes the context (large or small) in which we operate. Moreover, changes in context (e.g., the election of a maverick President) can change not just how factors interact with each other, but can introduce new factors and reconfigure the organization of factors in fundamental ways. This is a problem for machine learning, which operates on data sets that by definition were created previously, in a different context. Thus, taking context into account is a challenge for automation… “Our ability to manage and utilize emotion and to take into account the effects of context are key ingredients of critical thinking, creative problem solving, effective communication, adaptive learning, and good judgment. It has proven very difficult to program machines to emulate such human knowledge and skills, and it is not clear when (or whether) today’s fledgling efforts to do so will bear fruit.” |
“RCP Poll: K-12 Education Falls Short, and Hope for Gains Lags” on Real Clear Politics | K-12 education is a “social technology” whose improve performance (or not) will be a key driver of future labor productivity improvement. “One of a few areas that Democrats and Republicans can agree on these days are their views related to the state of education in the United States. It’s not good. Fifty-one percent of Republicans and 55% of Democrats rate American education as ‘only fair’ or ‘poor’”… Only 13% of respondents believe that in 20 years, the American public school system will be a model for excellence around the world. If the system can’t be changed, what are the chances it could be displaced by a new business model? AEI’s Rick Hess highlighted a hint of what might be possible in the future in a recent Education Week interview with Dan Ayoub, the the general manager of mixed reality, artificial intelligence, and STEM education for Microsoft. Intriguingly, before his current role Ayoub spent 20 years in the games industry, leading the development of the wildly popular game Halo. As Hess noted, “it often seems like game designers have figured out some things about engaging youths that have yet to show up in educational software. Ayoub’s reply was intriguing: “I think there’s a lot of what games do well that make sense in the classroom. Like making the student the center of the experience, gradually giving skills, and building on them. I think games are also great at teaching grit, resilience, and the understanding that failure is a part of success. Games are also increasingly social in nature, which is really interesting to think about in educational scenarios…” However, experience shows that simply adding advanced technology to a traditional business process never fully realizes the technology’s potential benefits. For that, you need to at minimum change the process, and often change the underlying business model. Unfortunately, both seem a long way off, at least in American K-12 education. |
| |
| |
| |
| |
| |
| |
| |
Aug19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Understanding Model Behavior Using Loops that Matter”, by Shoenberg et al | SURPRISE Systems Dynamics models complex non-linear behavior as a function of system structure, in particular interactions between system stocks (e.g., the level of water in a bathtub) and flows (e.g., the rates at which water is flowing into and out of the tub), and interactions between many positive and negative feedback loops. Up to now, discerning the dynamic impact of feedback loop interactions has been more art than science. This paper “presents a new and distinct method to find the 'loops that matter' in generating model behavior. This is a numeric method capable of determining the impact for every loop in a model and identifying which dominate behavior at each point in time.” This new development, in combination with automation and artificial intelligence technologies, should accelerate our ability to model and explain the operation of complex systems. However, SD is a top down modeling approach, unlike Agent Based Modeling, which is a bottom up method. As such, it cannot capture the evolving strategies of individual agents, and the way they drive the evolution of system structure. Still, this new method is still a significant advance. |
“Learning to Transfer Learn”, by Zhu et al from Google | SURPRISE Transfer learning – the ability to learn concepts in one context and apply them to significantly different contexts – is a fundamental aspect of fluid human intelligence that has thus far resisted replication via technology. This new paper highlights a new technique that has demonstrated a significant improvement in this area, which is critical to the continued development of artificial intelligence as a general purpose technology that can substantially increase productivity across the economy. |
“Competitive Multi-Agent Deep Reinforcement Learning with Counterfactual Thinking”, by Wang et al | The authors note that a core challenge in improving artificial intelligence methods in complex real world applications involving multiple agents is, “to find the optimized action policies for AI agents with limited knowledge about their environment… traditional single-agent reinforcement learning approaches ignore the interactions and the decision strategies of other competitors.” The authors describe a new method that enables agents to speed their learning about the complex environment by generating counterfactual strategies and estimating not the reward they would have produced, but the amount of regret for not having chosen them. They use deep reinforcement learning to optimize agent decisions by minimizing expected regret, mimicking, in some ways, the cognitive and emotional processes of (some) human decision makers. This is another example of how progress in relatively narrow technical areas is laying the groundwork for much faster improvement in integrated AI methods in the future. |
“Massive Multi-Agent Data-Driven Simulations of the GitHub Ecosystem”, by Blythe et al | This paper reports the results of one of the winning teams in a recent DARPA competition to use agent-based modeling and AI methods to accurately forecast the behavior of a very complex socio-technical system. It provides an excellent unclassified benchmark for the current state of development in this critical area. That said, there are two limitations to keep in mind while reading it. First, the forecast horizon was relatively short – one month. However, this involved the very complex interactions of three million agents. Second, the behaviors of most of these agents changed only slowly, unlike many other complex socio-technical systems (e.g., financial markets) in which behavior can change much faster, and is subject to non-linear jumps. |
“Playing a Strategy Game with Knowledge-Based Reinforcement Learning”, by Voss et al | This is another excellent benchmark paper, that focuses on the use of another key technology: the integration of existing knowledge (including causal knowledge) into reinforcement learning. In this case, the target environment is FreeCiv, which is a version of the complex game Civilization, which is familiar to many and is more similar to complex real world challenges than simpler games like chess. Reinforcement learning is used as a method to efficiently integrate the knowledge of multiple experts into strategy development, and to learn and increase that knowledge over time. Again, the potential future application to complex real world challenges is clear; the key issue, into which this paper provides insight, is how fast these methods are maturing and approaching the point at which they can be deployed in financial, commercial and other applications. |
“DeepMind’s Losses and the Future of Artificial Intelligence” by Gary Marcus | Gary Marcus has consistently questioned a lot of the hype that surrounds the current AI boom, and his writing is always thought-provoking. In his latest column, Marcus notes that, “DeepMind has been putting most of its eggs in one basket, a technique known as reinforcement learning. That technique combines deep learning, primarily for recognizing patterns, with reinforcement learning based on reward signals, such as a score in a game…The trouble is, the technique is very specific to a narrow range of circumstances…Current systems lack flexibility, and are unable to compensate if the world changes, sometimes even in tiny ways” (a point we have also often made). He also notes that an even more important issue is the difficulty, and often impossibility of explaining the results reached by deep learning algorithms (which are based on the very complex patterns it identifies in a data set), and thus the difficulty in persuading human decision makers to trust AI’s recommendations, particularly in an evolving environment. Marcus pointedly notes that because most challenging real world problems exist in complex, evolving environments, “DeepMind has yet to find any large-scale commercial application of deep reinforcement learning.” To his credit, Marcus also considers another conclusion: “it could simply be an issue of time” before commercial applications appear as AI technology improves. |
| |
| |
| |
| |
| |
| |
| |
Jul19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
The Technology Trap, by Carl Benedikt Frey | In his new book, Frey reminds us that the early stages of the industrial revolution was characterized by the mechanization of agriculture and manufacturing which substituted capital for labor and, while increasing productivity and returns to capital, led to widespread suffering, social disorder, and political conflict. It was only in the later stages of industrialization that capital goods became complements to labor, which augmented the latter’s productivity and wages (and which was itself dependent on substantial gains in education and the quality of human capital). Frey relates this to the exponential improvements in automation and AI technologies, and reminds us that we have been here before. |
Adoption of Automation Technologies: Evidence from Denmark, by Kromann and Sorensen | The authors provide relatively rare evidence about firm level adoption of automation technology. They find that it varies widely, with many slow adopters and a few very aggressive adopters, with many of the latter having high exposure to Chinese import competition. The authors also find that significant use of automation is associated with higher productivity growth and increases in profitability. Significantly, the authors conclude that, “the low use of automation to some extent is due to a lack of the necessary skills and resources to investigate the firms’ needs, possibilities for automation, and automation planning for the factory floor. The production managers were not unaware that automation technologies existed, but they were lacking knowledge or awareness regarding the specific technologies that they could invest in, on how to implement these, and on which production processes to automate.” |
“Robots or Workers? A Macro Analysis of Automation and Labor Markets” by Leduc and Liu from the Federal Reserve Bank of San Francisco | “Our model predicts that automation dampens wage growth, boosts labor productivity, and reduces the labor share of national income.” |
“The Metamorphosis” by Kissinger, Schmitdt, and Huttenlocher in The Atlantic | SURPRISE “Humanity is at the edge of a revolution driven by artificial intelligence…This revolution is unstoppable…We should accept that AI is bound to become increasingly sophisticated and ubiquitous, and ask ourselves: How will its evolution affect human perception, cognition, and interaction?” … “AI has enabled machines to play an increasingly decisive role in drawing conclusions from data and then taking action…The growing transfer of judgment from human beings to machines denotes the revolutionary aspect of AI…If AI improves constantly – and there is no reason to think it will not – the changes it will impose on human life will be transformative.” To illustrate this, the authors cite how the concept of nuclear deterrence – which is fundamentally based on a human desire to avoid self-destruction – could be radically transformed when AI plays a significant role in national security decision making. As they note, “the opacity and speed of the cyber world many overwhelm current planning models.” The authors conclude with this observation: “The challenge of absorbing this technology into the values and practices of the existing culture has no precedent. The most comparable event [in human history] was the transition from the medieval to the modern period.” |
| |
| |
| |
| |
| |
| |
| |
| |
| |
Jun19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
A New Law Suggests Quantum Supremacy Could Happen This Year” by Kevin Hartnett from Quanta Magazine, in Scientific American 21Jun19 | SURPRISE “Quantum computers are improving at a doubly exponential rate… That rapid improvement has led to what’s being called ‘Neven’s law,’ [named after Hartmut Neven Director of Google’s Quantum Artificial Intelligence Lab] a new kind of rule to describe how quickly quantum computers are gaining on classical ones… With double exponential growth, ‘it looks like nothing is happening, nothing is happening, and then whoops, suddenly you’re in a different world,’ Neven said. ‘That’s what we’re experiencing here.’” “The doubly exponential rate at which, according to Neven, quantum computers are gaining on classical ones is a result of two exponential factors combined with each other. “The first is that quantum computers have an intrinsic exponential advantage over classical ones: If a quantum circuit has four quantum bits, for example, it takes a classical circuit with 16 ordinary bits to achieve equivalent computational power. This would be true even if quantum technology never improved The second exponential factor comes from the rapid improvement of quantum processors. Neven says that Google’s best quantum chips have recently been improving at an exponential rate.” |
Progress in Quantum Computing creates the potential for an exponential increase in the speed at which artificial intelligence technologies improve. However, that also requires substantial advances in the rate at which software improves. As we have noted in past issues, this includes the rate at which AI software advances from associational/statistical approaches to ones based on far more difficult causal and counterfactual reasoning, which in turn heavily depend on advances in natural language processing. In June, four new papers provided new indications of progress in this area. | In “A Survey of Reinforcement Learning Informed by Natural Language”, Luketina et al find that, “To be successful in real-world tasks, Reinforcement Learning (RL) needs to exploit the compositional, relational, and hierarchical structure of the world, and learn to transfer it to the task at hand. Recent advances in representation learning for language make it possible to build models that acquire world knowledge from text corpora and integrate this knowledge into downstream decision making problems.” In “Shaping Belief States with Generative Environmental Models for Reinforcement Learning”, Gregor et al from DeepMind note that they “are interested in making agents that can solve a wide range of tasks in complex and dynamic environments. While tasks may be vastly different from each other, there is a large amount of structure in the world that can be captured and used by the agents in a task-independent manner. “This observation is consistent with the view that such general agents must understand the world around them. Algorithms that learn representations by exploiting structure in the data that are general enough to support a wide range of downstream tasks is what we refer to as unsupervised learning or self-supervised learning. We hypothesize that an ideal unsupervised learning algorithm should use past observations to create a stable representation of the environment. That is, a representation that captures the global factors of variation of the environment in a temporally coherent way. “As an example, consider an agent navigating in a complex landscape. At any given time, only a small part of the environment is observable from the perspective of the agent. The frames that this agent observes can vary significantly over time, even though the global structure oft he environment is relatively static with only a few moving objects. A useful representation of such an environment would contain, for example, a map describing the overall layout of the terrain. Our goal is to learn such representations in a general manner. Predictive models have long been hypothesized as a general mechanism to produce useful representations based on which an agent can perform a wide variety of tasks in partially observed worlds.” The authors then describe their progress toward achieving this goal, give an example of the current state of development, and describe the remaining obstacles to be overcome. In “Deep Reasoning Networks: Thinking Fast and Slow”, Chen et al address the challenge described by the DeepMind team, and “introduce Deep Reasoning Networks (DRNets), an end-to-end framework that combines deep learning with reasoning for solving complex tasks, typically in an unsupervised or weakly-supervised setting. DRNets exploit problem structure and prior knowledge by tightly combining logic and constraint reasoning with stochastic-gradient-based neural network optimization.” They conclude with examples that show their approach produces substantial gains in performance compared to previous techniques. Finally, in “Does It Make Sense? And Why? A Pilot Study for Sense Making and Explanation”, Wang et al observe that, “introducing common sense to natural language understanding systems has received increasing research attention. [Yet] It remains a fundamental question on how to evaluate whether a system has a sense making capability. Existing benchmarks measures commonsense knowledge indirectly and without explanation.” To accelerate this process, they “release a benchmark to directly test whether a system can differentiate natural language statements that make sense from those that do not make sense. In addition, a system is asked to identify the most crucial reason why a statement does not make sense. We evaluate models trained over large-scale language modeling tasks as well as human performance, and show the different challenges for system sense making.” |
"Predicting Research Trends with Semantic and Neural Networks with an Application in Quantum Physics", by Krenn and Zeilinger | This is a very thought-provoking paper on how advances in natural language processing and network analysis can be combined with a large body of textual data to both forecast and identify potential scientific (and engineering) advances. In theory, this approach has the potential to increase the productivity of R&D spending, which has been declining in recent years. |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
May19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Gifted classes may not help talented students move ahead faster”, by Jill Barshay, Hechinger Report | “A large survey of 2,000 elementary schools in three states found that not much advanced content is actually being taught to gifted students…“Teachers and educators are not super supportive of acceleration,” said Betsy McCoach, one of the researchers and a professor at the University of Connecticut.” At a time when performance is increasingly dependent on a small number of “hyperperformers” or “superstar” talent (e.g., see, The Best and The Rest: Revisiting The Norm Of Normality Of Individual Performance” by O’Boyle and Aguinis, and “Superstars: The Dynamics of Firms, Sectors, and Cities Leading the Global Economy” by Manyika et al from McKinsey), this finding that the education of America’s most talented students is largely being neglected by public schools has worrisome implications for future performance. |
“COBRA: Data-Efficient Model-Based RL through Unsupervised Object Discovery and Curiosity-Driven Exploration,” by Watters et al from Deep Mind | Progress in unsupervised reinforcement learning is a key indicator of developing AI capability. “Recent advances in deep reinforcement learning (RL) have shown remarkable success on challenging tasks. However, data efficiency and robustness to new contexts remain persistent challenges for deep RL algorithms, especially when the goal is for agents to learn practical tasks with limited supervision. Drawing inspiration from self-supervised “play” in human development, we introduce an agent that learns object-centric representations of its environment without supervision and subsequently harnesses these to learn policies efficiency and robustly.” |
“Cognitive Model Priors for Predicting Human Decisions” by Bourgin et al | SURPRISE This is yet another indicator of accelerating improvement in AI technologies. “Human decision-making underlies all economic behavior. For the past four decades, human decision-making under uncertainty has continued to be explained by theoretical models based on prospect theory, a framework that was awarded the Nobel Prize in Economic Sciences. However, theoretical models of this kind have developed slowly, and robust, high-precision predictive models of human decisions remain a challenge. “While machine learning is a natural candidate for solving these problems, it is currently unclear to what extent it can improve predictions obtained by current theories. We argue that this is mainly due to data scarcity, since noisy human behavior requires massive sample sizes to be accurately captured by off-the-shelf machine learning methods.” “To solve this problem, what is needed are machine learning models with appropriate inductive biases for capturing human behavior, and larger datasets. We offer two contributions towards this end: “First, we construct “cognitive model priors” by pretraining neural networks with synthetic data generated by cognitive models (i.e., theoretical models developed by cognitive psychologists). “We find that fine-tuning these networks on small datasets of real human decisions results in unprecedented state-of-the-art improvements on two benchmark datasets.” Second, we present the first large-scale dataset for human decision-making, containing over 240,000 human judgments across over 13,000 decision problems. This dataset reveals the circumstances where cognitive model priors are useful, and provides a new standard for benchmarking prediction of human decisions under uncertainty.” |
“Robots and Firms” by Koch et al | SURPRISE This study is based on unique micro-level evidence, and highlights the substantial disruption that lies ahead as the adoption of AI and automation technologies accelerates. The authors “study the implications of robot adoption at the level of individual firms using a rich panel data-set of Spanish manufacturing firms over a 27-year period (1990-2016). We focus on three central questions: (1) Which firms adopt robots? (2) What are the labor market effects of robot adoption at the firm level? (3) How does firm heterogeneity in robot adoption affect the industry equilibrium?... “As for the first question, we establish robust evidence that ex-ante larger and more productive firms are more likely to adopt robots, while ex-ante more skill-intensive firms are less likely to do so. As for the second question, we find that robot adoption generates substantial output gains in the vicinity of 20-25% within four years, reduces the labor cost share by 5-7% points, and leads to net job creation at a rate of 10%. Finally, we reveal substantial job losses in firms that do not adopt robots, and a productivity-enhancing reallocation of labor across firms, away from non-adopters, and toward adopters.” Unfortunately, the authors don’t report the impact on compensation of the employment changes they highlight. |
Coursera Global Skills Index 2019 | SURPRISE This report provides further evidence that exponentially improving automation and AI technologies seem increasingly likely to produce substantial economic and social disruption, with, at this point, uncertain by likely negative political consequences. “Two-thirds of the world’s population is falling behind in critical skills, including 90% of developing economies. Countries that rank in the lagging or emerging categories (the bottom two quartiles) in at least one domain [Business, Technology, and Data Science] make up 66% of the world’s population, indicating a critical need to upskill the global workforce. Such a large proportion of ill-prepared workers calls for greater investment in learning to ensure they remain competitive in the new economy… “Europe is the global skills leader. European countries make up over 80% of the cutting-edge category (top quartile globally) across Business, Technology, and Data Science. Finland, Switzerland, Austria, Sweden, Germany, Belgium, Norway, and the Netherlands are consistently cutting-edge in all the three domains. This advanced skill level is likely a result of Europe’s heavy institutional investment in education via workforce development and public education initiatives… “Skill performance within Europe still varies, though. Countries in Eastern Europe with less economic stability don’t perform as well as Western Europe in the three domains; Turkey, Ukraine, and Greece consistently land in the bottom half globally. “Asia Pacific, the Middle East and Africa, and Latin America have high skill inequality… “The United States must upskill while minding regional differences. Although known as a business leader for innovation, the U.S. hovers around the middle of the global rankings and is not cutting-edge in any of the three domains. While there’s a need for increased training across the U.S., skill levels vary between sub-regions. “The West leads in Technology and Data Science, reflecting the concentration of talent in areas like Silicon Valley. The Midwest shines in Business, ranking first or second in every competency except finance. “The South consistently ranks last in each domain and competency, suggesting a need for more robust training programs in the sub-region.” |
“The Wrong Kind of AI? Artificial Intelligence and the Future of Labor Demand” by Acemoglu and Restrepo | SURPRISE Written by two leading academic analysts of the economic and social impacts of advancing AI technologies and their implementation, this new paper provides an important warning about the disruption that lies ahead of current trends continue. “Artificial Intelligence is set to influence every aspect of our lives, not least the way production is organized. AI, as a technology platform, can automate tasks previously performed by labor or create new tasks and activities in which humans can be productively employed. “Recent technological change has been biased towards automation, with insufficient focus on creating new tasks where labor can be productively employed. The consequences of this choice have been stagnating labor demand, declining labor share in national income, rising inequality and lower productivity growth. “The current tendency is to develop AI in the direction of further automation, but this might mean missing out on the promise of the ‘right’ kind of AI with better economic and social outcomes.” |
| |
| |
| |
| |
| |
| |
| |
Apr19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Challenges of Real-World Reinforcement Learning”, by Dulac-Arnold et al | SURPRISE This paper is a timely reminder of the current limitations of this critical AI method when it applied to the complex, evolving problems that we frequently confront. “Reinforcement learning (RL) has proven its worth in a series of artificial domains, and is beginning to show some successes in real-world scenarios. However, much of the research advances in RL are often hard to leverage in real world systems due to a series of assumptions that are rarely satisfied in practice. “We present a set of nine unique challenges that must be addressed to ‘productionize’ RL to real world problems. At a high level, these challenges are: (1) Training off-line from the fixed logs of an external behavior policy. (2) Learning on the real system from limited samples. (3) High-dimensional continuous state and action spaces. (4) Safety constraints that should never or at least rarely be violated. (5) Tasks that may be partially observable, alternatively viewed as non-stationary or stochastic. (6) Reward functions that are unspecified, multi-objective, or risk-sensitive. (7) System operators who desire explainable policies and actions. (8) Inference that must happen in real-time at the control frequency of the system. (9) Large and/or unknown delays in the system actuators, sensors, or rewards. “While there has been research focusing on these challenges individually, there has been little research on algorithms that address all of these challenges together…An approach that does this would be applicable to a large number of real world problems.” |
“Survey on Automated Machine Learning”, by Zoller and Huber | Like the paper above, this is another important indicator of the type of obstacles that will need to be overcome in order to speed the deployment of artificial intelligence applications. “Machine learning has become a vital part in many aspects of our daily life. However, building well performing machine learning applications requires highly specialized data scientists and domain experts. Automated machine learning (AutoML) aims to reduce the demand for data scientists by enabling domain experts to automatically build machine learning applications without extensive knowledge of statistics and machine learning. In this survey, we summarize the recent developments in academy and industry regarding AutoML.” |
“Key Design Components and Considerations for Establishing a Single Payer Healthcare System”, by the US Congressional Budget Office | This is an excellent and brief summary of the key issues involved in establishing a more comprehensive single payer healthcare system in the United States. |
“Prices Paid to Hospitals by Private Health Plans are High Relative to Medicare [Prices] and Vary Widely” by White and Whaley from RAND | SURPRISE The US healthcare system is characterized by multiple prices being charged by healthcare suppliers depending on who is paying. The lowest prices are for beneficiaries of the federal Medicare system for senior citizens; the highest are for uninsured individuals. Based on a sample of hospital prices charged in 25 states, this new report from RAND finds that in some states, charges to private health insurance payers are more than 300% higher than those charged to Medicare. This translates into higher insurance premiums for employers and reduced take-home pay for employees as their benefits costs increase. The authors conclude that the potential for cost savings from a “single price” system are, frankly, enormous. Whether moving to a single payer system would be required to move to a single price system is an uncertainty, as it is not clear that private health insurance payers and employer who purchase such insurance have enough power to bring about this change on their own. |
“After 20 Years of Reform, Are America’s Schools Better Off?”, by Hess and Martin | In the short-term, lack of improvement in US education results will slow the deployment of AI and automation technologies; in the medium and long-term, however, it will speed many companies’ transition to “labor-lite” business models (e.g., based on AutoML technology noted above). This will likely worsen inequality and political conflict, increase social safety net spending, and lead to much higher taxes to fund it. “On the whole, it’s certainly possible to find some evidence of improvement — but progress is easiest to find in the metrics most amenable to manipulation…the U.S. also regularly administers the National Assessment of Educational Progress (NAEP) to a random, nationally representative set of schools. Because the NAEP isn’t linked to state accountability systems, it’s a good way to check the seemingly positive results of state tests. “From 2000 to 2017 (the most recent year for which data is available), NAEP scores showed that fourth-grade math results increased 14 points, which reflects a bit more than one year of extra learning. Eighth-grade math results also demonstrated significant improvement, increasing ten points in the same period. Fourth- and eighth-grade reading scores, meanwhile, barely budged. And almost all of the math gains were made in the decade from 2000 to 2010; performance has pretty much flatlined since then… “The Programme for International Student Assessment (PISA) is the only major international assessment of both reading and math performance. While PISA has its share of limitations, it offers a wholly independent view of American education and accountability systems. From the time PISA was first administered in 2000 to the most recent results from 2015, U.S. scores have actually declined, while America’s international ranking has remained largely static…there has been a lot of action, but not much in the way of demonstrated improvement. Just why this is the case remains an open question.” See also, “Why American Students Haven’t Gotten Better at Reading in 20 Years”, by Natalie Wexler in The Atlantic, and “US Achievement Gaps Hold Steady in the Face of Substantial Policy Initiatives” by Hanushek et al |
“Predicting Success in the Worldwide Start-Up Network”, by Nonaventura et al | This paper highlights the increasing ability of advanced AI techniques (combined, in this case, with social network analysis methods) to either automate or augment cognitively complex activities, while also improving predictive performance. The speed and breadth at which these emerging technologies will be deployed remains highly uncertain; however, this paper, and others like it, provide an indication of what lies ahead. “By drawing on large-scale online data we construct and analyze the time varying worldwide network of professional relationships among start-ups. The nodes of this network represent companies, while the links model the flow of employees and the associated transfer of know-how across companies. We use network centrality measures to assess, at an early stage, the likelihood of the long-term positive performance of a start-up, showing that the start-up network has predictive power and provides valuable recommendations doubling the current state of the art performance of venture funds. Our network-based approach not only offers an effective alternative to the labour-intensive screening processes of venture capital firms, but can also enable entrepreneurs and policy-makers to conduct a more objective assessment of the long-term potentials of innovation ecosystems and to target interventions accordingly.” |
| |
| |
| |
| |
| |
| |
| |
Mar19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Rigorous Agent Evaluation: An Adversarial Approach To Uncover Catastrophic Failures”, by Uesato et al from Deep Mind | Surprise This paper details how generative adversarial networks can be used to identify catastrophic failure modes in complex adaptive system. As this technology is further developed, it has potentially very important applications, in both the national security and economic sectors. |
“The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, And Sentences From Natural Supervision”, by Mao, et al from MIT, IBM, and DeepMind | Surprise The authors describe a potentially very powerful new approach to AI, which combines symbolic concept learning with a deep neural network. This leads to a substantial reduction in the amount of training data required, as well as a sharp improvement in the speed of concept learning, and thus, potentially, improvements in transfer learning (i.e., the application of concepts to novel situations). |
“Tackling Europe’s Gap in Digital and AI” by the McKinsey Global Institute | This report’s discouraging conclusions about the state of digitization and AI adoption in Europe has important implications for the region’s future economic growth, national security spending, social conditions, and political conflicts. “Digitisation is an important technical and organisational precondition for the spread of AI, yet Europe’s digital gap—at about 35 percent with the United States—has not narrowed in recent years. Early digital companies have been the first to develop strong positions in AI, yet only two European companies are in the worldwide digital top 30, and Europe is home to only 10 percent of the world’s digital unicorns...” “Europe has about 25 percent of AI startups, but its early-stage investment in AI lags behind that of the United States and China. Further, with the exception of smart robotics, Europe is not ahead of the United States in AI diffusion, and less than half of European firms have adopted one AI technology, with a majority of those still in the pilot stage.” |
“Informed Machine Learning – Towards a Taxonomy of Explicit Integration of Knowledge into Machine Learning”, by von Rueden et al | One of the constraints on the development and application of artificial intelligence technologies has been their need for large amounts of training data. We have previously noted the development of generative adversarial networks which generate artificial training data to speed the learning process. Another approach is the inclusion of domain knowledge to speed learning. This approach also potentially facilitates the development of abstractions by AI, which in turn facilitates “transfer learning”, or the application of conceptual abstractions to new situations, as is the case in human learning. This paper provides a useful taxonomy that helps you to understand this knowledge application process, and thus to develop different indicators as to its progress. |
“The Algorithmic Automation Problem: Prediction, Triage, and Human Effort”, by Rahu et al | As we have noted in past issues, the deployment of AI technologies is proceeding more slowly than some had expected. This paper describes how one of the underlying problems – determining whether an AI or a human being is best suited to perform a given task – can be more efficiently addressed. It thus provides another indicator that can be used to improve estimates of the speed of AI deployment, and thus their future impact on the economy, society, and politics. |
“How Aligned is Career and Technical Education to Local Labor Markets?”, by the Fordham Institute | Surprise As we have noted before, both education and healthcare provision are two critical enabling “social technologies” that have profound “downstream” impacts on economic, national security, social and political issues. This depressing report highlights the surprising extent to which CTE programs in the United States (usually known as Vocational and Technical Education in other nations) are not providing students with competences and credentials that are highly valued in private sector labor markets. If not corrected, this will contribute to weaker economic growth, more demand for social safety net spending, and probably higher levels of social and political conflict. |
“Causal Effect Identification from Multiple Incomplete Data Sources: A General Search-based Approach”, by Tikka et al | This is a very technical paper, but highlights use of Judea Pearl’s do-calculus in automated search for causal relationships in a large data set. As such, it is a key indicator of AI progress in the critical area of causal (and counterfactual) modeling. |
“How China tried and failed to win the AI race: The inside story”, by Alison Rayome in Tech Republic | Surprise The author claims that, “China fooled the world into believing it is winning the AI race, when really it is only just getting started.” However, a close reading of the article leaves us with a less smug conclusion, based on the distinction between a snapshot of a situation and how fast it is evolving over time. A more accurate conclusion seems to be that while the US is currently ahead of China in key areas of AI (including the chips on which AI software runs), in some areas the pace of improvement in China seems to be faster than the pace in the US (e.g., how lack of privacy concerns is enabling the creation of very large training data sets). In sum, the claim that “China has tried and failed to win the AI race” seems very premature. |
“Once Hailed as Unhackable, Blockchains are Getting Hacked” by Mike Orcutt in MIT Technology Review | Surprise “A blockchain is a cryptographic database maintained by a network of computers, each of which stores a copy of the most uptodate version. A blockchain protocol is a set of rules that dictate how the computers in the network, called nodes, should verify new transactions and add them to the database. The protocol employs cryptography, game theory, and economics to create incentives for the nodes to work toward securing the network instead of attacking it for personal gain. If set up correctly, this system can make it extremely difficult and expensive to add false transactions but relatively easy to verify valid ones.” “That’s what’s made the technology so appealing to many industries…But the more complex a blockchain system is, the more ways there are to make mistakes while setting it up.” “We’ve long known that just as blockchains have unique security features, they have unique vulnerabilities. Marketing slogans and headlines that called the technology “unhackable” were dead wrong. That’s been understood, at least in theory, since Bitcoin emerged a decade ago. But in the past year, amidst a Cambrian explosion of new cryptocurrency projects, we’ve started to see what this means in practice—and what these inherent weaknesses could mean for the future of blockchains and digital assets.” |
| |
| |
| |
| |
Feb19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“The Hanabi Challenge: A New Frontier for AI Research”, by Bard et All from DeepMind. | “From the early days of computing, games have been important testbeds for studying how well machines can do sophisticated decision making. In recent years, machine learning has made dramatic advances with artificial agents reaching superhuman performance in challenge domains like Go, Atari, and some variants of poker. As with their predecessors of chess, checkers, and backgammon, these game domains have driven research by providing sophisticated yet well-defined challenges for artificial intelligence practitioners. We continue this tradition by proposing the game of Hanabi as a new challenge domain with novel problems that arise from its combination of purely cooperative gameplay and imperfect information in a two to five player setting. In particular, we argue that Hanabi elevates reasoning about the beliefs and intentions of other agents to the foreground. We believe developing novel techniques capable of imbuing artificial agents with such theory of mind will not only be crucial for their success in Hanabi, but also in broader collaborative efforts, and especially those with human partners.” |
“Contest Models Highlight Inherent Inefficiencies Of Scientific Funding Competitions”, by Gross and Bergstrom | SURPRISE “Scientific research funding is allocated largely through a system of soliciting and ranking competitive grant proposals. In these competitions, the proposals themselves are not the deliverables that the funder seeks, but instead are used by the funder to screen for the most promising research ideas. Consequently, some of the funding program’s impact on science is squandered because applying researchers must spend time writing proposals instead of doing science. To what extent does the community’s aggregate investment in proposal preparation negate the scientific impact of the funding program? Are there alternative mechanisms for awarding funds that advance science more efficiently? We use the economic theory of contests to analyze how efficiently grant proposal competitions advance science, and compare them with recently proposed, partially randomized alternatives such as lotteries. “We find that the effort researchers waste in writing proposals may be comparable to the total scientific value of the research that the funding supports, especially when only a few proposals can be funded. Moreover, when professional pressures motivate investigators to seek funding for reasons that extend beyond the value of the proposed science (e.g., promotion, prestige), the entire program can actually hamper scientific progress when the number of awards is small. We suggest that lost efficiency may be restored either by partial lotteries for funding or by funding researchers based on past scientific success instead of proposals for future work.” |
“Bigger Teams Aren't Always Better in Science And Tech”, Science Daily, 13Feb19 and “The Challenge of Overcoming the “End of Science”: How can we improve R&D processes when they are so poorly defined?”, by Dr. Jeffrey Funk | Both of these articles provide further information about root causes of the apparent slowdown in the productivity of R&D. “In today's science and business worlds, it's increasingly common to hear that solving big problems requires a big team. But a new analysis of more than 65 million papers, patents and software projects found that smaller teams produce much more disruptive and innovative research…large teams, more often develop and consolidate existing knowledge.” Funk “discusses the falling productivity of R&D, the limitations of existing terms, concepts and theories of R&D, and the necessity of better defining R&D processes before the falling productivity of them can be reversed.” This is a very thought provoking paper that is well worth a read. |
Reverberations continue from EU regulators are taking aim at technology platform business models. | SURPRISE “The Google GDPR Fine: Some Thoughts”, by Michael Mandel “The GDPR will mean big changes in the way that European and U.S. companies do business in Europe. As we noted at a recent privacy panel, rather than being a matter of speculation, its economic impact has become an empirical question. Will the tighter privacy protections of the GDP slow growth and innovation, as skeptics claim, or will these provisions increase consumer trust and usher in a new era of European digital gains, as supporters say? We await the answers to these questions with great interest. “However, the enforcement stage of the GDPR has not gotten off on the right foot. CNIL, the French National Data Protection Commission, just fined Google 50 million euros for what they called ‘lack of transparency, inadequate information and lack of valid consent regarding the ads’ personalization.’ The fines were based in part on complaints filed by privacy groups on May 25, 2018, the very day that the GDPR went into effect. Moreover, the complaints were filed in France, despite that fact that Google’s European headquarters are in Ireland. “The location of the complaints is relevant because the most straightforward reading of the GDPR’s “one-stop-shop” principle suggests that the location of a company’s European headquarters is the main factor determining the company’s lead regulator for GDPR purposes. That’s not the only criterion, for sure, but it was only natural for the Irish Data Protection Commission to take the lead role in regulating Google. “The fact that the privacy organizations filed their complaints with France, not Ireland, suggests that they were forum-shopping–looking for a country which would look favorably on the issues they raised. Moreover, France’s willingness to jump to the front of the regulator queue suggests that they were interested in setting a precedent, rather than letting the GDPR process unfold. “Finally, an important part of the rationale behind the GDPR was to further move towards a digital single market, by allowing companies to only deal with a single privacy regulator. If other countries follow France’s lead and find reasons to levy data protection-related fines on multinationals that have their European headquarters elsewhere, then the GDPR will end up fragmenting markets, rather than making them more consistent. That’s a losing proposition for everyone.” “German Regulators Just Outlawed Facebook's Whole Ad Business”, Wired 7Feb19 “Germany’s Federal Cartel Office, the country’s antitrust regulator, ruled that Facebook was exploiting consumers by requiring them to agree to this kind of data collection in order to have an account, and has prohibited the practice going forward. Facebook has one month to appeal.” |
“Disinformation and Fake News: Final Report”, by the UK House of Commons | “We have always experienced propaganda and politically-aligned bias, which purports to be news, but this activity has taken on new forms and has been hugely magnified by information technology and the ubiquity of social media. In this environment, people are able to accept and give credence to information that reinforces their views, no matter how distorted or inaccurate, while dismissing content with which they do not agree as ‘fake news’. This has a polarising effect and reduces the common ground on which reasoned debate, based on objective facts, can take place. Much has been said about the coarsening of public debate, but when these factors are brought to bear directly in election campaigns then the very fabric of our democracy is threatened…. “The big tech companies must not be allowed to expand exponentially, without constraint or proper regulatory oversight. But only governments and the law are powerful enough to contain them. The legislative tools already exist. They must now be applied to digital activity, using tools such as privacy laws, data protection legislation, antitrust and competition law. If companies become monopolies they can be broken up, in whatever sector.” |
“Priority Challenges for Social and Behavioral Research and Its Modeling”, by Davis et al from RAND and “Uncertainty Analysis to Better Confront Model Uncertainty”, by Davis and Popper from RAND | SURPRISE “Social-behavioral (SB) modeling is famously hard. Three reasons merit pondering: First, Complex adaptive systems. Social systems are complex adaptive systems (CAS) that need to be modeled and analyzed accordingly—not with naïve efforts to achieve accurate and narrow predictions, but to achieve broader understanding, recognition of patterns and phases, limited forms of prediction, and results shown as a function of context and other assumptions. Great advances are needed in understanding the states of complex adaptive systems and their phase spaces and in recognizing both instabilities and opportunities for influence. Second, Wicked problems. Many social-behavioral issues arise in the form of so-called wicked problems— i. e., problems with no a priori solutions and with stakeholders that do not have stable objective functions. Solutions, if they are found at all, emerge from human interactions. Third, Structural dynamics. The very nature of social systems is often structurally dynamic in that structure changes may emerge after interactions and events. This complicates modeling… “The hard problems associated with CAS need not be impossible. It is not a pipe dream to imagine valuable SB modeling at individual, organizational, and societal scales. After all, complex adaptive systems are only chaotic in certain regions of their state spaces. Elsewhere a degree of prediction and influence is possible. We need to recognize when a social system is or is not controllable… As for problem wickedness, it should often be possible to understand SB phenomena well enough to guide actions that increase the likelihood of good developments and reduce the likelihood of bad ones. Consider how experienced negotiators can facilitate eventual agreements between nations, or between companies and unions, even when emotions run high and no agreement exists initially about endpoints. Experience helps, and model-based analysis can help to anticipate possibilities and design strategies. Given modern science and technology, opportunities for breakthroughs exist, but they will not come easily…. To improve SB modeling, we need to understand obstacles, beginning with shortcomings of the science that should underlie it. Current SB theories are many, rich, and informative, but also narrow and fragmented. They do not provide the basis for systemic SB modeling. More nearly comprehensive and coherent theories are needed, but current disciplinary norms and incentives favor continued narrowness and fragmentation. No ultimate “grand theory” is plausible, but a good deal of unification is possible with various domains… To represent social-behavioral theory requires increased emphasis on causal models (rather than statistical models) and on uncertainty-sensitive models that routinely display results parametrically in the multiple dimensions that define context. That is, we need models that help us with causal reasoning under uncertainty.” In the other RAND study on uncertainty, Davis and Popper note that, "the traditional focus of uncertainty analysis has been on model parameter uncertainty and irreducible variability (randomness), but as systems and problems have become more complex, uncertainty about the underlying conceptual model and its specification in code have become much more important. Better addressing model uncertainty is likely key to reducing policymaker skepticism about value of modeling results in decisionmaking. Their paper provides concrete recommendations for how to better address model related uncertainty. |
“Causal Effect Identification from Multiple Incomplete Data Sources: A General Search-based Approach”, by Tikka et al | This is a very technical paper, but highlights use of Judea Pearl’s do-calculus in automated search for causal relationships in a large data set. As such, it is a key indicator of AI progress in the critical area of causal (and counterfactual) modeling. |
“This Is Why AI Has Yet To Reshape Most Businesses” by Brian Bergstein in MIT Technology Review See also, “AI Adoption Advances, But Foundational Barriers Remain” by McKinsey | SURPRISE “Despite what you might hear about AI sweeping the world, people in a wide range of industries say the technology is tricky to deploy. It can be costly. And the initial payoff is often modest. It’s one thing to see breakthroughs in artificial intelligence that can outplay grandmasters of Go, or even to have devices that turn on music at your command. It’s another thing to use AI to make more than incremental changes in businesses that aren’t inherently digital… Gains have been largest at the biggest and richest companies, which can afford to spend heavily on the talent and technology infrastructure necessary to make AI work well…algorithms are a small part of what matters. Far more important are organizational elements that ripple from the IT department all the way to the front lines of a business…All this requires not just money but also patience, meticulousness, and other quintessentially human skills that too often are in short supply The McKinsey article notes that faster digitization is a critical enabler of faster AI deployment. |
“Companies Are Failing in Their Efforts to Become Data-Driven”, by Randy Bean and Thomas H. Davenport | “An eye-opening 77% of executives report that business adoption of Big Data/AI initiatives is a major challenge, up from 65% last year… “72% of survey participants report that they have yet to forge a data culture 69% report that they have not created a data-driven organization 53% state that they are not yet treating data as a business asset 52% admit that they are not competing on data and analytics.” “These sobering results and declines come in spite of increasing investment in big data and AI initiatives...Critical obstacles still must be overcome before companies begin to see meaningful benefits from their big data and AI investments... “Executives who responded to the survey say that the challenges to successful business adoption do not appear to stem from technology obstacles; only 7.5% of these executives cite technology as the challenge. Rather, 93% of respondents identify people and process issues as the obstacle. Clearly, the difficulty of cultural change has been dramatically underestimated in these leading companies — 40.3% identify lack of organization alignment and 24% cite cultural resistance as the leading factors contributing to this lack of business adoption.” |
OpenAI’s new Language Model is so powerful that its source code will not be released because of its potential for producing highly realistic fake news | From the OpenAI press release: “Natural language processing tasks, such as question answering, machine translation, reading comprehension, and summarization, are typically approached with supervised learning on task-specific datasets. We demonstrate that language models begin to learn these tasks without any explicit supervision when trained on a new dataset of millions of webpages called WebText.” For the full paper, see, “Language Models are Unsupervised Multitask Learners” by Radford et al Also, “Strategies for Structuring Story Generation” by Fan et al |
“World Discovery Models” by Azar et al from DeepMind | This paper is yet another indicator of the underappreciated speed at which various AI technologies and applications are developing. “The underlying process of discovery in humans is complex and multifaceted. However, one can identify two main mechanisms for discovery. The first mechanism is active information seeking. One of the primary behaviours of humans is their attraction to novelty (new information) in their world. The human mind is very good at distinguishing between the novel and the known, and this ability is partially due to the extensive internal reward mechanisms of surprise, curiosity and excitement The second mechanism is building a statistical world model. Within cognitive neuroscience, the theory of statistical predictive mind states that the brain, like scientists, constructs and maintains a set of hypotheses over its representation of the world. Upon perceiving a novelty, our brain has the ability to validate the existing hypothesis, reinforce the ones that are compatible with the new observation and discard the incompatible ones. This self-supervised process of hypothesis building is essentially how humans consolidate their ever-growing knowledge in the form of an accurate and global model…” “The outstanding ability of the human mind for discovery has led to many breakthroughs in science, art and technology. Here we investigate the possibility of building an agent capable of discovering its world using the modern AI technology… We introduce NDIGO, Neural Differential Information Gain Optimisation, a self-supervised discovery model that aims at seeking new information to construct a global view of its world from partial and noisy observations. Our experiments on some controlled 2-D navigation tasks show that NDIGO outperforms state-of-the-art information-seeking methods in terms of the quality of the learned representation. The improvement in performance is particularly significant in the presence of white or structured noise where other information-seeking methods follow the noise instead of discovering their world.” |
“Psychlab: A Psychology Laboratory for Deep Reinforcement Learning Agents”, by Letbo et al from DeepMind | This is a key indicator of advances in developing an AI based “theory of mind” that will enable artificial agents to better understand, and anticipate, the actions of human agents with whom they either partner or compete. “Psychlab is a simulated psychology laboratory inside the first-person 3D game world of DeepMind Lab. Psychlab enables implementations of classical laboratory psychological experiments so that they work with both human and artificial agents. Psychlab has a simple and flexible API that enables users to easily create their own tasks. As examples, we are releasing Psychlab implementations of several classical experimental paradigms including visual search, change detection, random dot motion discrimination, and multiple object tracking. We also contribute a study of the visual psychophysics of a specific state-of-the-art deep reinforcement learning agent: UNREAL. This study leads to the surprising conclusion that UNREAL learns more quickly about larger target stimuli than it does about smaller stimuli. In turn, this insight motivates a specific improvement in the form of a simple model of foveal vision that turns out to significantly boost UNREAL’s performance, both on Psychlab tasks, and on standard DeepMind Lab tasks. By open-sourcing Psychlab we hope to facilitate a range of future such studies that simultaneously advance deep reinforcement learning and improve its links with cognitive science.” |
“The Productivity Imperative For Healthcare Delivery In The United States”, by McKinsey | SURPRISE This new report examines how productivity in the US healthcare delivery industry evolved between 2001 and 2016. “There is little doubt that the trajectory of healthcare spending in the United States is worrisome and perhaps unsustainable. Underlying this spending is the complex system used to deliver healthcare services to patients. Given that the US currently expends 18% of its gross domestic product (GDP) on healthcare, this system might be expected to deliver high-quality, affordable, and convenient patient care—yet it often fails to achieve that goal… “One explanation, however, has largely been overlooked: poor productivity in the healthcare delivery industry. In practical terms, increased productivity in healthcare delivery would make it possible to continue driving medical advances and meet the growing demand for services while improving affordability (and likely maintaining current employment and wages)… Job creation—not labor productivity gains—was responsible for most of the growth in the US healthcare delivery industry from 2001 to 2016. Innovation, changes in business practices, and the other variables that typically constitute Multifactor Productivity Growth, harmed the industry’s growth. If the goal is to control healthcare spending growth, both trends must change… “The impact of improving productivity would be profound. Our conservative estimates suggest that if the healthcare delivery industry could rely more heavily on labor productivity gains rather than workforce expansion to meet demand growth, by 2028 healthcare spending could potentially be (on a nominal basis) about $280 billion to $550 billion less than current national health expenditures (NHE) projections suggest... Cumulatively, $1.2 trillion to $2.3 trillion could be saved over the next decade if healthcare delivery were to move to a productivity-driven growth model. Savings of this magnitude would bring the rise in healthcare spending in line with—and possibly below—GDP growth. In addition, the increased labor productivity in healthcare delivery would boost overall US economic growth at a faster rate than current projections—an incremental 20 to 40 basis points (bps) per annum—both through direct economic growth and the spillover impact of greater consumption in other industries. However, meaningful action by, and collaboration among, all stakeholders will be needed to deliver this value.” |
Jan19: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
DeepMind’s AlphaStar software has for the first time defeated a top ranked professional player five games to zero in the real time strategy game Starcraft II. | SURPRISE As DeepMind notes, “until now, AI techniques have struggled to cope with the complexity of StarCraft… The need to balance short and long-term goals and adapt to unexpected situations, poses a huge challenge…Mastering this problem required breakthroughs in several AI research challenges including: Game theory: StarCraft is a game where, just like rock-paper-scissors, there is no single best strategy. As such, an AI training process needs to continually explore and expand the frontiers of strategic knowledge. Imperfect Information: Unlike games like chess or Go where players see everything, crucial information is hidden from a StarCraft player and must be actively discovered by “scouting”. Long term planning: Like many real-world problems cause-and-effect is not instantaneous. Games can also take anywhere up to one hour to complete, meaning actions taken early in the game may not pay off for a long time. Real time: Unlike traditional board games where players alternate turns between subsequent moves, StarCraft players must perform actions continually as the game clock progresses. Large action space: Hundreds of different units and buildings must be controlled at once, in real-time, resulting in a huge combinatorial space of possibilities… Due to these immense challenges, StarCraft has emerged as a “grand challenge” for AI research.” DeepMind’s latest achievement is further evidence of the accelerating pace at which AI capabilities are improving. To be sure, Starcraft is still a discrete system, governed by an unchanging set of rules. In that sense, it critically differs from real world complex socio-technical systems, in which agents’ adaptive actions are not constrained by unchanging rules, and system dynamics evolve over time. In real world complex adaptive systems, making sense of new information in an evolving context, inducing abstract concepts from novel situations, and then using them to rapidly reason about the dynamics of a situation and the likely impact of possible actions remain, for now, beyond the capabilities of the most advanced artificial intelligence systems. In addition, some critics have noted that DeepStar’s victory owed more to the speed of its play relative to its very talented human opponent, and physical accuracy of its moves (placement of units on a map) than it did to superior strategy (e.g., see, “An Analysis On How Deepmind’s Starcraft 2 AI’s Superhuman Speed is Probably a Band-Aid Fix For The Limitations of Imitation Learning”, by Aleksi Pietikainen). But all that said, DeepStar’s success still reminds us that the gap between the capabilities of AI and human beings, even in cognitively challenging areas, is closing faster than many people appreciate |
“Quantum Terrorism” Collective Vulnerability of Global Quantum Systems” by Johnson et al. | SURPRISE Quantum computing, while exponentially more powerful than today’s technology, will also bring new vulnerabilities. The authors “show that an entirely new form of threat arises by which a group of 3 or more quantum-enabled adversaries can maximally disrupt the global quantum state of future systems in a way that is practically impossible to detect, and that is amplified by the way that humans naturally group into adversarial entities.” |
Shoshana Zuboff’s new book, “The Age of Surveillance Capitalism: The Fight for the Future at the New Frontier of Power” crystallizes the often unspoken worries that many people have felt about exponentially improving artificial intelligence technologies. | SURPRISE Writing in the Financial Times Zuboff describes “a new economic logic [she] calls ‘surveillance capitalism’”. It “was invented in the teeth of the dot.com bust, when a fledgling company called Google decided to try and boost ad revenue by using its exclusive access to largely ignored data logs — the “digital exhaust” left over from users’ online search and browsing. The data would be analysed for predictive patterns that could match ads and users. Google would both repurpose the “surplus” behavioural data and develop methods to aggressively seek new sources of it…These operations were designed to bypass user awareness and, therefore, eliminate any possible “friction”. In other words, from the very start Google’s breakthrough depended upon a one-way mirror: surveillance…” Surveillance capitalism soon migrated to Facebook and rose to become the default model for capital accumulation in Silicon Valley, embraced by every start-up and app. It was rationalised as a quid pro quo for free services but is no more limited to that context than mass production was limited to the fabrication of the Model T. It is now present across a wide range of sectors, including insurance, retail, healthcare, finance, entertainment, education and more. Capitalism is literally shifting under our gaze.” “It has long been understood that capitalism evolves by claiming things that exist outside of the market dynamic and turning them into market commodities for sale and purchase. Surveillance capitalism extends this pattern by declaring private human experience as free raw material that can be computed and fashioned into behavioural predictions for production and exchange…” “Surveillance capitalists produce deeply anti-democratic asymmetries of knowledge and the power that accrues to knowledge. They know everything about us, while their operations are designed to be unknowable to us. They predict our futures and configure our behaviour, but for the sake of others’ goals and financial gain. This power to know and modify human behaviour is unprecedented. “Often confused with totalitarianism and feared as Big Brother, it is a new species of modern power that I call ‘instrumentarianism’. [This] power can know and modify the behaviour of individuals, groups and populations in the service of surveillance capital. The Cambridge Analytica scandal revealed how, with the right knowhow, these methods of instrumentarian power can pivot to political objectives. But make no mistake, every tactic employed by Cambridge Analytica was part of surveillance capitalism’s routine operations of behavioural influence.” As the Guardian noted in its review of her book, Zuboff, “points out that while most of us think that we are dealing merely with algorithmic inscrutability, in fact what confronts us is the latest phase in capitalism’s long evolution – from the making of products, to mass production, to managerial capitalism, to services, to financial capitalism, and now to the exploitation of behavioural predictions covertly derived from the surveillance of users.” “The combination of state surveillance and its capitalist counterpart means that digital technology is separating the citizens in all societies into two groups: the watchers (invisible, unknown and unaccountable) and the watched. This has profound consequences for democracy because asymmetry of knowledge translates into asymmetries of power. But whereas most democratic societies have at least some degree of oversight of state surveillance, we currently have almost no regulatory oversight of its privatised counterpart. This is intolerable.” |
In light of Zuboff’s book, the provocatively titled article (“The French Fine Against Google is the Start of a War”) in the 24Jan19 Economist does not seem excessive. | SURPRISE “On January 21st France’s data-protection regulator, which is known by its French acronym, CNIL, announced that it had found Google’s data-collection practices to be in breach of the European Union’s new privacy law, the General Data Protection Regulation (GDPR). CNIL hit Google with a €50m ($57m), the biggest yet levied under GDPR. Google’s fault, said the regulator, had been its failure to be clear and transparent when gathering data from users…” “The fine represents the first volley fired by European regulators at the heart of the business model on which Google and many other online services are based, one which revolves around the frictionless collection of personal data about customers to create personalised advertising. It is the first time that the data practices behind Google’s advertising business, and thus those of a whole industry, have been deemed illegal. Google says it will appeal against the ruling. Its argument will not be over whether consent is required to collect personal data—it agrees that it is—but what quality of consent counts as sufficient…Up to now the rules that underpin the digital economy have been written by Google, Facebook et al. But with this week’s fine that is starting to change.” The growing public anger in the West over reduced privacy that both Zuboff’s book and the CNIL fine represent has important implications for the race to create ever more powerful machine learning/artificial intelligence capabilities, whose advancement is critically dependent on access to large amounts of training data. In China, data privacy is not an issue. In Europe, it is a very serious issue today. The US currently lies somewhere in between. While emerging technologies like Generative Adversarial Networks may in future be used to quickly generate high quality simulated data that can be used to train AI, we aren’t there yet. Until we are, the data privacy issue will be inextricably linked to the pace of AI development, which in turn has national security, as well as economic and social implications. |
“We analyzed 16,625 papers to figure out where AI is headed next” by Karen Hao, in MIT Technology Review, 25Jan19 | “The sudden rise and fall of different techniques has characterized AI research for a long time, he says. Every decade has seen a heated competition between different ideas. Then, once in a while, a switch flips, and everyone in the community converges on a specific one. At MIT Technology Review, we wanted to visualize these fits and starts. So we turned to one of the largest open-source databases of scientific papers, known as the Arxiv (pronounced “archive”). We downloaded the abstracts of all 16,625 papers available in the “artificial intelligence” section through November 18, 2018, and tracked the words mentioned through the years to see how the field has evolved…” ”We found three major trends: a shift toward machine learning during the late 1990s and early 2000s, a rise in the popularity of neural networks beginning in the early 2010s, and growth in reinforcement learning in the past few years…The biggest shift we found was a transition away from knowledge-based systems by the early 2000s. These computer programs are based on the idea that you can use rules to encode all human knowledge. In their place, researchers turned to machine learning—the parent category of algorithms that includes deep learning…Instead of requiring people to manually encode hundreds of thousands of rules, this approach programs machines to extract those rules automatically from a pile of data. Just like that, the field abandoned knowledge-based systems and turned to refining machine learning…” “In the few years since the rise of deep learning, our analysis reveals, a third and final shift has taken place in AI research. As well as the different techniques in machine learning, there are three different types: supervised, unsupervised, and reinforcement learning. Supervised learning, which involves feeding a machine labeled data, is the most commonly used and also has the most practical applications by far. In the last few years, however, reinforcement learning, which mimics the process of training animals through punishments and rewards, has seen a rapid uptick of mentions in paper abstracts… [The pivotal] moment came in October 2015, when DeepMind’s AlphaGo, trained with reinforcement learning, defeated the world champion in the ancient game of Go. The effect on the research community was immediate… Our analysis provides only the most recent snapshot of the competition among ideas that characterizes AI research. But it illustrates the fickleness of the quest to duplicate intelligence… Every decade, in other words, has essentially seen the reign of a different technique: neural networks in the late ’50s and ’60s, various symbolic approaches in the ’70s, knowledge-based systems in the ’80s, Bayesian networks in the ’90s, support vector machines in the ’00s, and neural networks again in the ’10s. The 2020s should be no different, meaning the era of deep learning may soon come to an end.” |
“It’s Still the Prices, Stupid”, by Anderson et all in Health Affairs | As we have noted, healthcare and education are critical “social technologies”, particularly in a period of rapid change and heightened uncertainty about employment (which, for many Americans, is the source of their health insurance). Improving the effectiveness, efficiency, and adaptability of both these technologies will have a critical impact on the economy, society, and politics in the future. The authors of this article update a famous 2003 article titled “It’s the Prices, Stupid”, which “found that the sizable differences in health spending between the US and other countries were explained mainly by health care prices.” The authors of the present article find that, “The conclusion that prices are the primary reason why the US spends more on health care than any other country remains valid, despite health policy reforms and health systems restructuring that have occurred in the US and other industrialized countries since the 2003 article’s publication. On key measures of health care resources per capita (hospital beds, physicians, and nurses), the US still provides significantly fewer resources compared to the OECD median country. Since the US is not consuming greater resources than other countries, the most logical factor is the higher prices paid in the US.” |
On the education front, Colorado recently updated its “Talent Pipeline” Report, which provides a stark reminder of how poorly the US education system is performing, even in a state with the nation’s second highest percentage of residents with a bachelors degree or higher (about 40%). | Out of 100 students who complete 9th grade, 70 graduate from high school on time, 43 enroll in college that autumn, 32 return after their first year of college, and just 25 graduate from college within six years of starting it. Results like these have two critical implications. First, they point to stagnant or declining human capital, which is a key driver of total factor productivity and thus long-term growth, particularly as the economy becomes more knowledge intensive. Second, at a time when the capabilities of labor substituting technologies (like AI and automation) have been improving exponentially, the failure of human capital to keep pace will naturally induce businesses to invest more in the former and less in the latter, which would likely produce worsening economic inequality, rising unemployment, and skyrocketing government spending on social safety net programs – which will have to be paid for wither with higher taxes or unlikely cuts in other spending. |
| |
| |
Dec18: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Profiling for IQ Opens Up New Uber-Parenting Possibilities”, Financial Times, 22Nov18 | SURPRISE “A US start-up, Genomic Prediction, claims it can genetically profile embryos to predict IQ, as well as height and disease risk. Since fertility treatment often produces multiple viable embryos, only one or two of which can be implanted, prospective parents could pick those with the “best” genes.” The FT notes the implications of this technology development: “We are sliding into Gattaca territory, in which successive generations are selected not only for health but also for beauty, intellect, stature and other aptitudes. Parents may even think it their moral duty to choose the “best” possible baby, not just for themselves but to serve the national interest. Carl Shulman and Nick Bostrom, from the Future of Humanity Institute at Oxford university, predicted in 2013 that some nations may pursue the idea of a perfectible population to gain economic advantage. China, incidentally, has long been reading the genomes of its cleverest students.” If this technology is scaled up, the economic, national security, social, and political implications will be both profound and highly disruptive. |
“Natural Language Understanding Poised to Transform How We Work”, Financial Times, 3Dec19 | The FT notes, “If language understanding can be automated in a wide range of contexts, it is likely to have a profound effect on many professional jobs. Communication using written words plays a central part in many people’s working lives. But it will become a less exclusively human task if machines learn how to extract meaning from text and churn out reports.” Up to now, an obstacle “in training neural networks to do the type of work analysts face — distilling information from several sources — is the scarcity of appropriate data to train the systems. It would require public data sets that include both source documents and a final synthesis, giving a complete picture that the system could learn from. [However], despite challenges such as these, recent leaps in natural language understanding (NLU) have made these systems more effective and brought the technology to a point where it is starting to find its way into many more business applications.” The article focuses on technology from Primer (www.primer.ai), a new AI start-up, that has developed more capable natural language understanding technology that is now in use by intelligence agencies. |
“Why Companies that Wait to Adopt AI May Never Catch Up”, Harvard Business Review, by Mahidhar and Davenport, 3Dec18 | SURPRISE If the authors’ hypothesis is correct, then AI will lead to the intensification of “winner take all” markets, with more companies and business models struggling to earn economic profits. Research has shown that this will also lead to worsening income inequality between employees at the winning companies and everyone else. |
“Your Smartphone’s AI Algorithms Could Tell if You are Depressed”, MIT Technology Review, 3Dec18 | Reports on a Stanford study (“Measuring Depression Symptom Severity from Spoken Langage and 3D Facial Expresisons” by Haque et al) that used “a combination of facial expressions, voice tone, and use of specific words was used to diagnose depression”, with 80% accuracy. This could well turn out to be a two edged sword, with clear benefits for early diagnosis and treatment of mental illness, but equally important concerns about privacy (e.g., its use by employers or insurance companies). |
The Artificial Intelligence Index 2018 Annual Report | This report provides a range of excellent benchmarks for measuring the rate of improvement for various AI technologies, and their adoption across industries. Key findings include substantial shortening of training times (e.g., for visual recognition tasks), and the increasing rate of improvement for natural language understanding based on the GLUE benchmark. |
“Learning from the Experts: From Expert Systems to Machine Learned Diagnosis Models” by Ravuri et al | This paper describes how a model that embodies the knowledge of domain experts was used to generate artificial (synthetic) data about a system that was then used to train a deep learning network. This is an interesting approach that bears monitoring, particularly its potential future application to agent based modeling of complex adaptive systems. |
“How Artificial Intelligence will Reshape the Global Order: The Coming Competition Between Digital Authoritarianism and Liberal Democracy”, by Nicholas Wright | SURPRISE A thought provoking forecast of how developing social control technologies could affect domestic politics and strengthen authoritarian governments. |
“Data Breaches Could Cause Users to Opt Out of Sharing Personal Data. Then What?” by Douglas Yeung from RAND | “If the public broadly opts out of using tech tools…insufficient or unreliable user data could destabilize the data aggregation business model that powers much of the tech industry. Developers of technologies such as artificial intelligence, as well as businesses built on big data, could not longer count on ever-expanding streams of data. Without this data, machine learning models would be less accurate.” Arguably, with its General Data Protection Regulation (GDPR), the European Union has already moved in this direction. While the author focuses on commercial issues, there are also national security implications if China – where data privacy is not recognized as a legitimate concern – is able to develop superior AI applications because of access to a richer set of training data. |
“Parents 2018: Going Beyond Good Grades”, a report by Learning Heroes and Edge Research | SURPRISE Improving education, and more broadly the quality of a nation’s human capital, is critical to improving employment, productivity and economic growth and reducing income inequality. But no system, team, or individual can improve (except by random luck) in the absence of accurate feedback. And this new report makes painfully clear that this is too often missing in America’s K-12 education system. The report begins with the observation that, “parents have high aspirations for their children. Eight in 10 parents think it’s important for their child to earn a college degree, with African-American and Hispanic parents more likely to think it’s absolutely essential or very important. Yet if students are not meeting grade-level expectations, parents’ aspirations and students’ goals for themselves are unlikely to be realized. Today, nearly 40% of college students take at least one remedial course; those who do are much more likely to drop out, dashing both their and their parents’ hopes for the future… Over three years, one alarming finding has remained constant: Nearly 9 in 10 parents, regardless of race, income, geography, and education levels, believe their child is achieving at or above grade level. Yet national data indicates only about one-third of students actually perform at that level. In 8th grade mathematics, while 44% of white students scored at the proficient level on the National Assessment of Educational Progress in 2017, only 20% of Hispanic and 13% of African-American students did so. This year, we delved into the drivers of this “disconnect.” We wanted to understand why parents with children in grades 3-8 hold such a rosy picture of their children’s performance and what could be done to move them toward a more complete and accurate view… Report Cards Sit at the Center of the Disconnect: Parents rely heavily on report card grades as their primary source of information and assume good grades mean their child is performing at grade level. Yet two-thirds of teachers say report cards also reflect effort, progress, and participation in class, not just mastery of grade-level content… More than 6 in 10 parents report that their child receives mostly A’s and B’s on their report card, with 84% of parents assuming this indicates their child is doing the work expected of them at their current grade… Yet a recent study by TNTP found that while nearly two-thirds of students across five school systems earned A’s and B’s, far fewer met grade-level expectations on state tests. On the whole, students who were earning B’s in math and English language arts had less than a 35% chance of having met the grade-level bar on state exams.” |
Nov18: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Deep Learning can Replicate Adaptive Traders in a Limit-Order Book Financial Market”, by Calvez and Cliff | “Successful human traders, and advanced automated algorithmic trading systems, learn from experience and adapt over time as market conditions change…We report successful results from using deep learning neural networks (DLNNs) to learn, purely by observation, the behavior of profitable traders in an electronic market… We also demonstrate that DLNNs can learn to perform better (i.e., more profitably) than the trader that provided the training data. We believe that this is the first ever demonstration that DLNNs can successfully replicate a human-like, or super-human, adaptive trader.” This is a significant development. Along with similar advances in reinforcement learning (e.g., by Deep Mind with AlphaZero), one can easily envision a situation where – at least over short time frames – most humans completely lose their edge over algorithms. The good news (form humans at least) is that over longer time frames, the structure of the system evolves (and becomes less discrete), and performance becomes more dependent on higher forms of reasoning – causal and counterfactual – where humans are still far ahead of algorithms (and whose sensemaking, situation awareness, and decision making The Index Investor is intended to support ). |
“Social media cluster dynamics create resilient global hate highways”, by Johnson et al | “Online social media allows individuals to cluster around common interests -- including hate. We show that tight-knit social clusters interlink to form resilient ‘global hate highways’ that bridge independent social network platforms, countries, languages and ideologies, and can quickly self-repair and rewire. We provide a mathematical theory that reveals a hidden resilience in the global axis of hate; explains a likely ineffectiveness of current control methods; and offers improvements…” |
“The Semiconductor Industry and the Power of Globalization” The Economist 1Dec18 | “If data are the new oil...chips are what turn them into something useful.” This special report provides a good overview of how the critical and highly globalized semiconductor supply chain is coming under increased pressure as competition between China and the United States intensifies. |
“There’s a Reason Why Teachers Don’t Use the Software Provided by Their Districts”, by Thomas Arnett | SURPRISE We have noted in the past that education (like healthcare) is a critical social technology where substantial performance improvement is critical to increasing future rates of national productivity growth and reducing inequality. This study is not encouraging with respect to the impact technology has been having on the education sector. The authors find that, “a median of 70% of districts’ software licenses never get used, and a median of 97.6% of licenses are never used intensively. |
Reports emerged from China that gene editing CRISPR technology to modify a human embryo’s DNA before implanting it in a woman’s womb via IVF. The initial focus was reportedly on producing children who are resistant to HIV, smallpox, and cholera. | SURPRISE While this has been recognized as a possibility, there was also a belief that it would not happen so quickly, or with so little control. It was also significant that the target of the DNA modification was resistance to smallpox, a disease which is believed to have been eradicated and whose causal agents are now only retained by governments (which makes them potentially very powerful biowar weapons). |
“Virtual Social Science” by Sefan Thurner. | SURPRISE Thurner is one of the world’s leading complex adaptive systems researchers, and anything he writes is usually rich with unique insights. His latest paper is no exception. He reviews findings from the analysis of 14 years of extremely rich data from Pardus, a massive multiplayer online game (MMOG) involving about 430,000 players in which economic, social, and other decisions are made by humans, not algorithms. This data can be used to develop and test a wide range of social science theories about the behavior of complex adaptive systems at various levels of aggregation, from the individual to the group to the system. It can also be used to evaluate agent-based, and AI driven approaches to predicting the future behavior of complex systems The author shows how many of the findings from analyzing game data line up with experimental findings based on the behavior of far fewer subjects. This points the way towards a new and potentially much more powerful approach to social science. However, Thurner also notes the current limits on the extent to which human societies can be understood, and their behavior predicted using this methodology: the inherent “co-evolutionary complexity” of complex adaptive social system, whose interactions cause structures to change over time, often in a non-linear manner. |
Oct18: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
“Using Machine Learning to Replicate Chaotic Attractors”, by Pathak et al | SURPRISE. Advances in a machine learning area known as “reservoir computing” have led to the creation of a model that reproduced the dynamics of a complex dynamical system. If this initial work can be extended it represents a significant advance. That said, this is not the same thing as AI learning and being able to reproduce and predict the dynamic behavior of a complex adaptive system, such as financial markets, and economies. |
“The Impact of Bots on Opinions in Social Networks” by Hjouji et al | Using both a model and data from the 2016 US presidential election, the authors conclude that “a small number of highly active bots in a social network can have a disproportionate impact on opinion…due to the fact that bots post one hundred times more frequently than humans.” In theory, this should make it easier for platforms like Twitter and Facebook to identify and close down these bots. The authors also surprisingly found that in 2016 pro-Clinton bots produced opinion shifts that were almost twice as large as the pro-Trump bots, despite the latter being larger in number. |
“Learning-Adjusted Years of Schooling” by Filmer et al from the World Bank | This valuable new indicator metric combines both the time spent in school and how much is learned during that time. The authors find that LAYS is strongly correlated with GDP growth. They also find wide gaps between countries, with some education systems being much more productive (in terms of learning per unit of time) than others. The good news is that this points to a substantial source of future gains for these economies in total factor productivity, provided their education systems can be improved. |
“The Condition of College and Career Readiness, 2018” by ACT Inc. | More disappointing results based on a well-known indicator of US K-12 education system performance. About three-fourths (76%) of 2018 ACT-tested graduates said they aspire to postsecondary education. Most of those students said they aspire to a four-year degree or higher. Only 27% met all four C&C ready benchmarks; 35% met none. Readiness levels in math have steadily declined since 2014. Sample size = 1.9m. “Just 26% of ACT-tested 2018 graduates likely have the foundational work readiness skills needed for more than nine out of 10 jobs recently profiled in the ACT JobPro® database. This has significant (and negative) implications for future productivity and wage growth. |
Sep18: New Technology Information: Indicators and Surprises | Why Is This Information Valuable? |
In a new book, “AI Superpowers”, Chinese venture capitalist Kai-Fu Lee makes an important point: There is a critical difference between AI innovation and AI implementation. Success in the latter depends on the ability to collect and analyze large amounts of data – and this is an area where China is outpacing the rest of the world, because of its size, its state capitalism model, low level of concern with privacy, and its data intensive approach to domestic security. | Provides a logical argument for how and why China could gain a significant advantage in key artificial intelligence technologies. |
US House of Representatives Subcommittee on Information Technology published a new report titled “Rise of the Machines.” Highlights: “First, AI is an immature technology; its abilities in many areas are still relatively new. Second, the workforce is affected by AI; whether that effect is positive, negative, or neutral remains to be seen. Third, AI requires massive amounts of data, which may invade privacy or perpetuate bias, even when using data for good purposes. Finally, AI has the potential to disrupt every sector of society in both anticipated and unanticipated ways.” | Report’s conclusions are an interesting contrast to Kai-Fu Lee’s. Similar to critiques by Gary Marcus and Judea Pearl, it highlights the limitations of current AI technologies, which suggests we are further away from a critical threshold than many media reports would suggest. That said, it also agrees that once a critical threshold of AI capability is reached, it will have strong disruptive effects. However, the report agrees with Lee that privacy concerns are a potentially important constraint on AI progress. |
“China is Overtaking the US in Scientific Research” by Peter Orzag in Bloomberg Opinion. Not just the quantity, but also “the quality of Chinese research is improving, though it currently remains below that of U.S. academics. A recent analysis suggests that, measured not just by numbers of papers but also by citations from other academics, Chinese scholars could become the global leaders in the near future.” | Suggests that the pace of technological improvement in China will accelerate. |
“Quantum Hegemony: China’s Ambitions and the Challenge to US Innovation Leadership”. Center for a New American Security. “China’s advances in quantum science could impact the future military and strategic balance, perhaps even leapfrogging traditional U.S. military-technological advantages. Although it is difficult to predict the trajectories and timeframes for their realization, these dual-use quantum technologies could “offset” key pillars of U.S. military power, potentially undermining critical technological advantages associated with today’s information-centric ways of war, epitomized by the U.S. model.” | Highlights a key area in which faster Chinese technological progress and breakthroughs could confer substantial military advantage. |
“A Storm in an IoT Cup: The Emergence of Cyber-Physical Social Machines” by Madaan et al. “The concept of ‘social machines’ is increasingly being used to characterize various socio-cognitive spaces on the Web. Social machines are human collectives using networked digital technology, which initiate real-world processes and activities including human communication, interactions and knowledge creation. As such, they continuously emerge and fade on the Web. The relationship between humans and machines is made more complex by the adoption of Internet of Things (IoT) sensors and devices. The scale, automation, continuous sensing, and actuation capabilities of these devices add an extra dimension to the relationship between humans and machines making it difficult to understand their evolution at either the systemic or the conceptual level. This article describes these new socio-technical systems, which we term Cyber-Physical Social Machines.” | Increasing complexity creates exponentially more hidden critical thresholds, and ways for a system to generate non-linear effects. |
“Notes From the Frontier: Modeling the Impact of AI on the World Economy”, McKinsey Global Institute. Adoption of AI could increase annual global GDP growth by 1.2%. Adoption of AI technologies and emergence of their impact is following typical “S-Curve” pattern. At this point, “the absence of evidence is not evidence of absence” of its potential impact. | Excellent analysis of the current state of AI development, rate of adoption, and range of observed effects. Critical Point: “Because economic gains combine and compound over time…a key challenge is that adoption of AI could widen gaps between countries, companies, and workers.” |
“Blueprint: How DNA Makes Us Who We Are” by Robert Plomin. Argues that genetic differences cause most variation in human psychological traits. Accumulating evidence for the dominance of nature over nurture has many potentially disruptive implications. See also, "Top 10 Replicated Findings from Behavioral Genetics" by Plomin et al | Surprise. The implications of the body of research this book compiles and synthesizes has enormous disruptive potential, at the economic, social, and ultimately political level. |