スピーカー紹介

Josh Tenenbaum WEBSITE

Massachusetts Institute of Technology

Building Machines That See, Think and Learn Like People

Recent advances in deep learning with neural networks have transformed both AI technologies and computational neuroscience, and sparked new excitement in connecting the science and engineering of intelligence. But while deep neural networks excel primarily at pattern recognition and function approximation, human intelligence goes much deeper. I will talk about prospects and approaches for reverse-engineering intelligence based on the idea of modeling the world: the computations, algorithms, data structures and neural mechanisms that enable our minds to explain and understand what we see, or imagine things we have never seen; to set goals, make plans and solve problems as we try to make those real; and to build new models as we learn more about the world, learning from both our successes as well as our failures.

I will focus on human capacities for common-sense scene understanding — perceiving the world in terms of physical objects, intentional agents, and their causal interactions — and learning the generative models that support these inferences. I will introduce some of the technical concepts based on probabilistic programs, simulation engines for intuitive physics and intuitive psychology, and Bayesian program induction, that we use to model these cognitive capacities, and some of the experimental methods we use to test these models behaviorally. I will also talk about how these models can connect to and build on the recent neural network revolution: the role of deep learning systems in making inference and learning more efficient, as part of an integrated neuro-symbolic-probabilistic modeling and inference toolkit, and first steps in relating the models we have built to brain systems in human and non-human primates

Biosketch

Josh Tenenbaum is Professor of Computational Cognitive Science in the Department of Brain and Cognitive Sciences at MIT, and a member of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and Center for Brains, Minds and Machines (CBMM). He received his PhD from MIT in 1999, and was an Assistant Professor at Stanford University from 1999 to 2002 before returning to MIT. His long-term goal is to reverse-engineer how intelligence arises in the human mind and brain, and draw on these insights to engineer more human-like machine intelligence. In cognitive science, he is best known for developing accounts of cognition as probabilistic inference in structured generative models, and applying this approach to the study of human concept learning, causal reasoning, language acquisition, visual perception, intuitive physics, and theory of mind. In AI, he and his group have developed widely used models for nonlinear dimensionality reduction, probabilistic programming, and Bayesian unsupervised learning and structure discovery. His current work focuses on commonsense visual scene understanding and its neural basis, the development of common sense in infants and young children, and models of learning as program induction. He and his students’ work have received best paper or best student paper awards or honorable mentions at leading conferences in Cognitive Science, Computer Vision, Neural Information Processing Systems, Reinforcement Learning and Decision Making, Robotics, and the Society for Philosophy and Psychology. He is the recipient of the Distinguished Scientific Award for Early Career Contribution to Psychology from the American Psychological Association (2008), the Troland Research Award from the National Academy of Sciences (2012), the Howard Crosby Warren Medal from the Society of Experimental Psychologists (2015), the R&D Magazine Innovator of the Year (2018), and a MacArthur Fellowship (2019). He is a fellow of the Cognitive Science Society, the Society for Experimental Psychologists, and a member of the American Academy of Arts and Sciences.
Yann LeCun WEBSITE

Facebook AI Research & New York University

Self-Supervised Learning

What learning paradigm do humans and animal use? The dominant machine learning paradigms, supervised and reinforcement learning, require many more labeled samples and many more trials to learn a task that animals and humans. The brain seems capable of learning a model of how the world works, in a task-independent way, largely through observation. Prediction using this model can be used for planning and reasoning. World models accumulate large amounts of background knowledge about the world which, perhaps, constitute the basis of common sense.

One promising avenue is self-supervised learning (SSL), where the machine predicts parts of its input from other parts of its input. SSL has already brought about great progress in discrete domains, such as language modeling and understanding. The question is how to use SSL for high-dimensional continuous domains such as audio, images and video. I will review the most promising current approaches within the conceptual framework of Energy-based Models.

Biosketch

Yann LeCun is VP and Chief AI Scientist at Facebook and Silver Professor at NYU affiliated with the Courant Institute and the Center for Data Science. He was the founding Director of Facebook AI Research and of the NYU Center for Data Science. He received an EE Diploma from ESIEE (Paris) in 1983, a PhD in Computer Science from Sorbonne Université (Paris) in 1987. After a postdoc at the University of Toronto, he joined AT&T Bell Laboratories. He became head of the Image Processing Research Department at AT&T Labs-Research in 1996, and joined NYU in 2003 after a short tenure at the NEC Research Institute. In late 2013, LeCun became Director of AI Research at Facebook, while remaining on the NYU Faculty part-time. He was visiting professor at Collège de France in 2016. His research interests include machine learning and artificial intelligence, with applications to computer vision, natural language understanding, robotics, and computational neuroscience. He is best known for his work in deep learning and the invention of the convolutional network method which is widely used for image, video and speech recognition. He is a member of the US National Academy of Engineering, a Chevalier de la Légion d’Honneur, a fellow of AAAI, the recipient of the 2014 IEEE Neural Network Pioneer Award, the 2015 IEEE Pattern Analysis and Machine Intelligence Distinguished Researcher Award, the 2016 Lovie Award for Lifetime Achievement, the University of Pennsylvania Pender Award, and honorary doctorates from IPN, Mexico and EPFL. He is the recipient of the 2018 ACM Turing Award (with Geoffrey Hinton and Yoshua Bengio) for “conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing”.
Yutaka Matsuo WEBSITE

The University of Tokyo

松尾豊

東京大学

World Model for Perception, Control and Language

World models are unsupervised/self-supervised predictive models of how the world evolves. In this talk, I will give our recent advances on deep generative models for building a model of complex worlds. Specifically, I introduce multi-modal variational autoencoders (VAEs), for learning joint embedding of multi-modal observations, e.g., image and text, and its several applications. I also introduce our recent advances on amortized inference, for handling a set of observations and achieving flexible and efficient posterior inference. Finally, I will discuss how the world models play an important role for developing efficient robot control and “true” natural language understanding, and hypothetical architectures for combining world models and components of natural language processing.

Biosketch

Yutaka Matsuo is a professor at Graduate School of Engineering, the University of Tokyo. He received his BS, MS, and Ph.D. degrees from the University of Tokyo in 1997, 1999, and 2002. After working at National Institute of Advanced Industrial Science and Technology (AIST) and Stanford University, he joined the faculty of University of Tokyo in 2007. He served as Editor-in-chief from 2012 to 2014, and as the chair of the ELSI committee from 2014 to 2018 at Japan Society for Artificial Intelligence (JSAI). He is now the president of Japanese Deep Learning Association (JDLA), and a member of the board of director at SoftBank Group Corp.

He is working on artificial intelligence, especially on deep learning and web mining.
Doina Precup WEBSITE

McGill University

Fast Reinforcement Learning with Generalized Policy Updates

The combination of reinforcement learning with deep learning is a promising approach to tackle important sequential decision-making problems that are currently intractable. One obstacle to overcome is the amount of data needed by learning systems of this type. In this talk, I will describe an approach to this problem which relies on decomposing complex problems into multiple tasks that unfold in sequence or in parallel. By associating each task with a reward function, this problem decomposition can be seamlessly accommodated within the standard reinforcement learning formalism. This can be done through a generalization of two fundamental operations in reinforcement learning: policy improvement and policy evaluation. The generalized version of these operations allow one to leverage the solution of some tasks to speed up the solution of other tasks. I will describe computational approaches for implementing these operations, as well as their known connections to behavioral science as well as neuroscience (through successor features). Computational experiments illustrate the power of our approach.

Biosketch

Doina Precup splits her time between McGill University, where she co-directs the Reasoning and Learning Lab in the School of Computer Science, and DeepMind Montreal, where she has led the research team since its formation in October 2017. Her research interests are in the areas of reinforcement learning, deep learning, time series analysis, and diverse applications of machine learning in health care, automated control, and other fields. She became a senior member of the Association for the Advancement of Artificial Intelligence in 2015, Canada Research Chair in Machine Learning in 2016, Senior Fellow of the Canadian Institute for Advanced Research in 2017, and received a Canada CIFAR AI (CCAI) Chair in 2018. She currently serves as the Associate Scientific Director of the Healthy Brains, Healthy Lives research effort at McGill. Dr. Precup is also involved in activities supporting the organization of Mila and the wider Montreal and Quebec AI ecosystem.
David Silver WEBSITE

DeepMind

Deep Reinforcement Learning from AlphaGo to AlphaStar

Could the objective of maximising reward be enough to drive intelligent behaviour?

Reward-maximising systems have achieved remarkable success in several challenging problems for artificial intelligence, by combining reinforcement learning with deep neural networks. In this talk I describe the ideas and algorithms that led to AlphaGo: the first program to defeat a human champion in the game of Go; AlphaZero: which learned, from scratch, to also defeat the world computer champions in chess and shogi; and AlphaStar: the first program to defeat a human champion in the real-time strategy game of StarCraft.

Biosketch

David Silver is a principal research scientist at DeepMind and a professor at University College London. David’s work focuses on artificially intelligent agents based on reinforcement learning. David co-led the project that combined deep learning and reinforcement learning to play Atari games directly from pixels (Nature 2015). He also led the AlphaGo project, culminating in the first program to defeat a top professional player in the full-size game of Go (Nature 2016), and the AlphaZero project, which learned by itself to defeat the world’s strongest chess, shogi and Go programs (Nature 2017, Science 2018). Most recently he co-led the AlphaStar project, which led to the world’s first grandmaster level StarCraft player (Nature 2019). His work has been recognised by the Marvin Minsky award, Mensa Foundation Prize, and Royal Academy of Engineering Silver Medal.
Masashi Sugiyama WEBSITE

RIKEN Center for Advanced Intelligence Project
The University of Tokyo International Research Center for Neurointelligence

杉山将

理化学研究所革新知能統合研究センター
東京大学国際高等研究所ニューロインテリジェンス国際研究機構

Recent Advances in Robust Machine Learning

In real-world machine learning applications, robustness to various factors such as sensor noise, human error, insufficient information, adversarial attack, changing environments, and sample selection bias is becoming increasingly important. In this talk, I will give an overview of our recent advances in robust machine learning, including noise transition estimation, noisy sample removal, weakly supervised classification, domain adaptation, distributionally robust learning, and classification with rejection.

Biosketch

Masashi Sugiyama received a PhD degree in Computer Science from Tokyo Institute of Technology, Japan in 2001. He has been a Professor at the University of Tokyo since 2014, and concurrently he has been the inaugural director of RIKEN Center for Advanced Intelligence Project since 2016. He (co-)authored machine learning monographs/textbooks such as Machine Learning in Non-Stationary Environments (MIT Press 2012), Density Ratio Estimation in Machine Learning (Cambridge University Press 2012), Statistical Reinforcement Learning (Chapman and Hall/CRC 2015), Introduction to Statistical Machine Learning (Morgan Kaufmann 2015), and Variational Bayesian Learning Theory (Cambridge University Press 2019). He served as a Program Co-chairs for NeuriPS2015 and AISTATS2019. He received the Japan Academy Medal in 2017.
Ila Fiete WEBSITE

Massachusetts Institute of Technology

Mixed Modular Codes and Remapping for Highly Generalizable Learning and Inference

It has long been hypothesized that generalizably solving complex problems involves decomposing them into simpler components and combining these parts in effective ways to solve new instances. How does the brain do so? The hippocampal complex is a rich playground for understanding how the brain constructs cognitive representations of latent variables because of its quantifiable representations of spatial information navigation. I will describe how the hippocampal complex — consisting of the hippocampus and associated areas – factorizes the world by generating separate low-dimensional manifolds to integrate and store different metric (Euclidean) variables relevant for navigation. I will show that these manifolds are rigid, persisting unchanged across tasks and states including waking and sleep. Next, I will describe how the same simple and rigid low-dimensional recurrent circuits could represent variables of different dimension and topology, with high capacity. Such generalization can be achieved by combining rigid modular representations with feedforward learning of inputs and random output projections to a sparsely active layer. These models require no rewiring of the local recurrent circuitry and thus permit relatively fast learning; the high capacity may enable lifelong learning. I will conclude by discussing how combining local metric maps with the dynamics of hippocampal remapping could permit learning of and inference on complex knowledge graphs. These models shed light on how the same neural circuits could play a role in both spatial and non-spatial cognitive mapping and inference.

Biosketch

Ila Fiete is a Professor in the Department of Brain and Cognitive Sciences and the McGovern Institute at MIT. Her group seeks to understand why the brain contains particular codes, how the architecture and dynamics of neural circuits shape such codes, and how neural circuit dynamics perform desired computations that unfold over time. They are specifically interested in how the brain learns, holds memories, integrates, and performs cognitive inference and reasoning. They use analytical and computational tools, and their approach includes working closely with collaborators on specific experimental systems.
Ila Fiete obtained her Ph.D. at Harvard, under the guidance of Sebastian Seung (then at MIT). Her postdoctoral work was at the Kavli Institute for Theoretical Physics at Santa Barbara, and at Caltech, where she was a Broad Fellow. She subsequently joined the faculty of the University of Texas at Austin in the Center for Learning and Memory . Ila Fiete is an HHMI Faculty Scholar. She has been a CIFAR Senior Fellow, a McKnight Scholar, an ONR Young Investigator, an Alfred P. Sloan Foundation Fellow and a Searle Scholar
Karl Friston WEBSITE

University College London

Active Inference and Artificial Curiosity

This talk offers a formal account of insight and learning in terms of active (Bayesian) inference. It deals with the dual problem of inferring states of the world and learning its statistical structure. In contrast to current trends in machine learning (e.g., deep learning), we focus on how agents learn from a small number of ambiguous outcomes to form insight. I will use simulations of abstract rule-learning and approximate Bayesian inference to show that minimising (expected) free energy leads to active sampling of novel contingencies. This epistemic, curiosity-directed behaviour closes `explanatory gaps’ in knowledge about the causal structure of the world; thereby reducing ignorance, in addition to resolving uncertainty about states of the known world. We then move from inference to model selection or structure learning to show how abductive processes emerge when agents test plausible hypotheses about symmetries in their generative models of the world. The ensuing Bayesian model reduction evokes mechanisms associated with sleep and has all the hallmarks of aha moments.

Biosketch

Karl Friston is a theoretical neuroscientist and authority on brain imaging. He invented statistical parametric mapping (SPM), voxel-based morphometry (VBM) and dynamic causal modelling (DCM). These contributions were motivated by schizophrenia research and theoretical studies of value-learning, formulated as the dysconnection hypothesis of schizophrenia. Mathematical contributions include variational Laplacian procedures and generalized filtering for hierarchical Bayesian model inversion. Friston currently works on models of functional integration in the human brain and the principles that underlie neuronal interactions. His main contribution to theoretical neurobiology is a free-energy principle for action and perception (active inference). Friston received the first Young Investigators Award in Human Brain Mapping (1996) and was elected a Fellow of the Academy of Medical Sciences (1999). In 2000 he was President of the international Organization of Human Brain Mapping. In 2003 he was awarded the Minerva Golden Brain Award and was elected a Fellow of the Royal Society in 2006. In 2008 he received a Medal, College de France and an Honorary Doctorate from the University of York in 2011. He became of Fellow of the Royal Society of Biology in 2012, received the Weldon Memorial prize and Medal in 2013 for contributions to mathematical biology and was elected as a member of EMBO (excellence in the life sciences) in 2014 and the Academia Europaea in (2015). He was the 2016 recipient of the Charles Branch Award for unparalleled breakthroughs in Brain Research and the Glass Brain Award, a lifetime achievement award in the field of human brain mapping. He holds Honorary Doctorates from the University of Zurich and Radboud University.
Yukie Nagai WEBSITE

The University of Tokyo International Research Center for Neurointelligence

長井志江

東京大学国際高等研究所ニューロインテリジェンス国際研究機構

Cognitive Development Based on Predictive Coding

A theoretical framework called predictive coding suggests that the human brain works as a predictive machine. That is, the brain tries to minimize prediction errors by updating the internal model and/or by affecting the environment. We have been investigating to what extent the predictive coding theory accounts for human cognitive development in terms of the temporal continuity (i.e., from non-social to social development) and the individual diversity (i.e., differences between typical development and developmental disorders).

This talk presents computational neural networks we designed to examine whether and how the process of minimizing prediction errors lead to cognitive development. Our experiments using a robot demonstrated that various cognitive abilities such as goal-directed action, imitation, estimation of others’ intention, and altruistic behavior emerged in a staged manner as observed in infants. Not only the characteristics of typical development but also those of developmental disorders such as autism spectrum disorder were generated as a result of aberrant prediction abilities. These results demonstrate that predictive coding provides a unified computational account for cognitive development (Nagai, Phil Trans B 2019).

Biosketch

Yukie Nagai is a Project Professor at International Research Center for Neurointelligence, the University of Tokyo. She received her Ph.D. in Engineering from Osaka University in 2004 and worked as a Postdoc Researcher at National Institute of Information and Communications Technology (NICT) and at Bielefeld University for five years. She then became a Specially Appointed Associate Professor at Osaka University in 2009 and a Senior Researcher at NICT in 2017. Since April 2019, she is working with the University of Tokyo.

Yukie Nagai has been investigating underlying neural mechanisms for social cognitive development by means of computational approaches. She designs neural network models for robots to learn to acquire cognitive functions such as self-other cognition, estimation of others’ intention and emotion, altruism, and so on. Her developmental theory based on predictive coding accounts for both continuity and diversity of cognitive development. She serves as the research director of JST CREST Cognitive Mirroring since December 2016.
Maneesh Sahani WEBSITE

Gatsby Computational Neuroscience Unit

How Do Neural Systems Learn to Infer?

How neural systems learn to integrate noisy and incomplete information to enable an organism to perceive, choose and move effectively in a range of different environments, with little supervision or reinforcement, remains poorly understood. Behavioural experiments provide evidence to support an optimal strategy defined by probabilistic or Bayesian computation; and neurally plausible schemes for the representation of uncertainty and for basic computations have been proposed. However, a fully integrated theory of representation, computation and learning has yet to achieve close agreement with experimental findings. In this talk I will argue that the widespread architectural motif of interconnected recurrent circuits is ideally suited to implement a form of inferential computation in a way that accords with known neural responses, and that processes of adaptive and associative learning within such architectures may serve to learn accurate probabilistic models of sensory experience. The resulting mathematical framework offers a novel way to relate artificial learning systems to neural circuits, while avoiding the challenges of credit assignment that accompany task-specific models.

Biosketch

Maneesh Sahani is Professor of Theoretical Neuroscience and Machine Learning and Director of the Gatsby Computational Neuroscience Unit at University College London (UCL). He received his Ph.D. in 1999 from the Computation and Neural Systems program at Caltech under the supervision of Richard Andersen and John Hopfield. After postdoctoral work in the Gatsby Unit and at UCSF, he joined the faculty at Gatsby in 2004, becoming Director in 2017. His work spans the interface of the fields of machine learning and neuroscience, with particular emphasis on the computations underlying inference and control under uncertainty, in perceptual and motor cortical systems.
Tadahiro Taniguchi WEBSITE

Ritsumeikan University

谷口忠大

立命館大学

Symbol Emergence in Robotics: Pursuing Integrative Cognitive Architecture using Probabilistic Generative Models for Real-world Language Acquisition

Symbol emergence in robotics aims to develop a robot that can adapt to the real-world environment, human linguistic communications, and acquire language from sensorimotor information alone, i.e., in an unsupervised manner. This line of studies is essential not only for creating a robot that can collaborate with people through human-robot interactions but also for understanding human cognitive development. This invited lecture introduces the recent development of integrative probabilistic generative models for language learning, e.g., spatial concept formation with simultaneous localization and mapping, and vision of symbol emergence in robotics. I will also introduce challenges related to the integration of probabilistic generative models and deep learning for integrative language learning by robots.

Biosketch

Tadahiro Taniguchi received the ME and Ph.D. degrees from Kyoto University, in 2003 and 2006, respectively. From April 2005 to March 2006, he was a Japan Society for the Promotion of Science (JSPS) Research Fellow (DC2) at the Department of Mechanical Engineering and Science, Graduate School of Engineering, Kyoto University. From April 2006 to March 2007, he was a JSPS Research Fellow (PD) at the same department. From April 2007 to March 2008, he was a JSPS Research Fellow at the Department of Systems Science, Graduate School of Informatics, Kyoto University. From April 2008 to March 2010, he was an Assistant Professor at the Department of Human and Computer Intelligence, Ritsumeikan University. From April 2010 to March 2017, he was an Associate Professor at the same department. From September 2015 to September 2016, he is a Visiting Associate Professor at the Department of Electrical and Electronic Engineering, Imperial College London.

From April 2017, he has been a Professor at the Department of Information Science and Engineering, Ritsumeikan University. From April 2017, he has been a visiting general chief scientist, AI solution center, Panasonic, as well. He has been engaged in research on machine learning, emergent systems, intelligent vehicle and symbol emergence in robotics.
Matthew Botvinick WEBSITE

DeepMind

Object-oriented deep learning

Recent progress in deep learning indicates that neural networks may be capable of structured reasoning in compositional domains, an ability that has been long argued to demand quite different machinery. The most widely discussed results have been in the domain of language modeling. I will describe recent research from our group extending the same approach to visual cognition. We know from neuro- and cognitive science that the human visual system carves visual scenes into discrete objects, and represents these objects in terms of their relations to one another. Building from this insight, we investigate deep learning systems that, likewise, process visual data in terms of objects and relations. Endowing deep networks with this inductive bias yields state-of-the-art performance in prediction and interpretation of visual events.

Biosketch

Director of Neuroscience Research, DeepMind, and Honorary Professor, Gatsby Computational Neuroscience Unit, University College London, 2016-present. B.A., Stanford, 1989; M.D., Cornell, 1994; Ph.D. in Cognitive and Computational Neuroscience, Carnegie Mellon University, 2001; Assistant Professor of Psychiatry, University of Pennsylvania, 2002-2007; Professor of Psychology and Neuroscience, Princeton, 2007-2016. Dr. Botvinick’s research focuses on issues spanning artificial intelligence, neuroscience and cognitive science, with a focus on goal-directed behavior and reinforcement learning. At DeepMind, he leads an interdisciplinary team pursuing research that straddles the boundary between neuroscience and AI.
Ryota Kanai WEBSITE

ARAYA Inc.

金井良太

株式会社アラヤ

Consciousness and Intelligence

In this talk, I will present our current two lines of theoretical work on consciousness. In the first half, I will discuss our recent work on the information closure theory of consciousness (ICT; Chang et al., 2020) in which we argue that the scale and the boundary of consciousness are determined by the so-called information closure of the system. In the ICT, we introduce the notion of non-trivial information closure (NTIC) to explain how the content of consciousness can be informationally closed from the environment and at the same time mirrors the events in the external world. We discuss our preliminary results on how Bayesian agents may develop NTIC over a short time period and how NTIC may be disrupted by prediction errors.

In the second half, I will present a hypothesis that consciousness has evolved as a platform for general intelligence (Kanai et al. in preparation). Here we define general intelligence as the ability to apply knowledge and models acquired from past experiences to generate solutions to novel problems. We propose three ways to establish general intelligence in AI systems, namely solution by simulation, solution by combination and solution by generation. Then, we relate those solutions to putative functions of consciousness put forward, respectively, by the information generation hypothesis (Kanai et al., 2019), the global workspace theory, and a form of higher order theory where qualia are regarded as meta-representations. By linking theories of consciousness to possible architectures of artificial general intelligence, I will discuss the fundamental link between consciousness and intelligence.

References

Chang, A.Y., Biehl, M., Yu, Y., & Kanai, R. (2020). Information Closure Theory of Consciousness. Front. Psychol. doi: 10.3389/fpsyg.2020.01504 Kanai, R., Chang, A., Yu, Y., Magrans de Abril, I., Biehl, M., & Guttenberg, N. (2019) Information generation as a functional basis of consciousness, Neuroscience of Consciousness, Volume 2019, Issue 1, 2019, niz016, https://doi.org/10.1093/nc/niz016

Biosketch

Ryota Kanai is the founder and CEO of Araya, Inc. His research focusses on understanding the computational basis of consciousness. He studies consciousness via multi-disciplinary approaches combining neuroscience, information theory, and artificial intelligence. He received his Ph.D from Utrecht University on visual perception in 2005 and continued his research on consciousness at California Institute of Technology and University College London. In 2012, he joined Sackler Centre for Consciousness Science, University of Sussex as a Reader in Cognitive Neuroscience. Then he returned to his native Japan and founded Araya, Inc. with the ambition to scale up his research programme via the mechanism of startups. He and his team at Araya work towards discovering functional and computational principles underlying biological consciousness.
Angela Langdon WEBSITE

Princeton University

Model-based Reward Prediction: Algorithms for Learning to Represent a Task

Reinforcement learning provides a robust and powerful suite of algorithms for forming predictions about future outcomes based solely on experience in an environment. Learning these reward predictions fundamentally rely on the abstract concept of a ‘state’—an internal representation used by an agent, whether animal, human or artificial, to summarize the features of the external and internal environment that are relevant for guiding behavior on a particular task. Armed with this summary representation, an agent can make decisions, perform actions and learn accurate reward predictions in order to interact effectively with the world. However, it remains an open question how a useful and efficient state representation might be constructed from continuous experience.

Phasic activity in dopamine neurons has long been identified as a neural correlate of reward prediction error signals in the brain. Using recent results on the properties of dopamine neurons during learning, I will show how these prediction error signals reflect more dimensions of an expected outcome than scalar reward value. These features imply a richer learning process in the brain than what is typically assumed in simple RL models. By relaxing the fundamental assumptions of model-free reinforcement learning, and revisiting the concept of a state, I will introduce an algorithmic framework in which reward predictions are embedded in learned expectations about the structure of the task environment—a form of ‘model-based’ reinforcement learning in which continuous experience can be parsed into a flexible and compact state representation. Using this framework, I will highlight a number of intriguing and perhaps surprising implications for the algorithmic understanding of learning in both biological and artificial agents.

Biosketch

Angela Langdon is an Associate Research Scholar at Princeton University, co-directing the Niv Lab at the Princeton Neuroscience Institute together with Yael Niv. Her research focuses on the neural computations that support reward prediction and learning in the brain.

She brings tools from her background in physics, PhD training in computational neuroscience and postdoctoral research in reinforcement learning to develop biologically plausible algorithms that give novel insight into how animals, humans and artificial agents learn to predict events and outcomes directly from observing and interacting with their environment.
Hiroyuki Nakahara WEBSITE

RIKEN Center for Brain Science

中原裕之

理化学研究所　脳神経科学研究センター

Neural Computations for Making Decisions with Others’ Rewards and Decisions

In social life, humans need to make decisions, considering others. In particular, they often make decisions, taking account of others’ rewards, and also frequently have to consider others’ decisions to make better decisions. These two capability is fundamental for our social behavior. Here, we present two human fMRI studies for investigating neural underpinning for decision-making with others’ rewards as well as with prediction of others’ decisions. In the first study, we investigated how a key social signal, or value of reward to others, is converted into self-oriented decision making. We show three-stage processing of social value conversion from the offer to the effective value and then to the final decision value, mediated from the right temporoparietal junction (rTPJ) and the left dorsolateral prefrontal cortex (ldlPFC) to the right anterior insula (rAI) and then to the medial prefrontal cortex (mPFC), respectively. Further, we found that these characteristics of social conversion underlie distinct sociobehavioral phenotypes, namely between prosocial and selfish individuals. In the second study, we investigated how a decision is supported by prediction of others’ decision in the face of its uncertainty. We show a complementary processing of others’ decision predictions when the prediction is easy or difficult; neural signals of one’s own decisions are modulated by those of the prediction on other’s decisions, in the ventral striatum by amygdala for the easy case, and the dorsomedial prefrontal cortex by the right superior temporal sulcus (rSTS) for the difficult case. These findings indicate a large network of neural systems for social behavior in the human brain.

Biosketch

Hiroyuki Nakahara is a principle investigator at RIKEN Center for Brain Science, and adjunct professor at Kyoto University. He received Ph.D. from University of Tokyo on his studies on sequential decision-making in biological systems. His laboratory is interested in understanding computational principles that underlie the way neural systems realize adaptive behavior, decision-making, and associated learning. In particular, we focus on (1) reward-based learning and decision-making and (2) social learning and decision-making. We often combine our theoretical studies with human fMRI experiment. We also work to develop advanced methods of analysis, for instance, using techniques of information geometry.
Xiao-Jing Wang WEBSITE

New York University

Learning to Learn and the Brain

Two-way interplay between neuroscience and AI represents a fertile research ground; neuroscientists increasingly use tools from machine learning in building neural network models, and AI researchers continue to be inspired from brain research on topics such as computational role of feedback projections and importance of executive control by the prefrontal cortex. One open question for both fields is learning-to-learn, i.e. how learning benefits from previous learning experiences. To explore this question mechanistically, we trained a recurrent neural network (RNN) model on a series of delayed sensorimotor association problems. In each problem, a pair of visual stimuli had to be uniquely mapped to a pair of actions. The stimuli were represented as random vectors in high-dimensional feature space and changed from problem to problem, while the actions stayed fixed. Training was performed in an online manner to mimic behavioral training in animals. First, we show that RNNs trained in this manner on a series of problems exhibit learning-to-learn despite the absence of an explicit meta-learning objective: early problems were learned in hundreds of trials, while later problems were learned within twenty trials on average. Second, dimensionality reduction on the population activity showed that the learned representations exhibit a low-dimensional structure that encoded decisions and choices, and a higher-dimensional representation of the stimulus. Further, the latent decision representations were shared across problems despite plasticity-driven changes to the population activity that ensued from learning. Control simulations indicated that the formation of these latent representations is necessary for rapid learning, suggesting that the network extracts the latent task structure and reuses it to learn new problems more efficiently. Third, we find that the number of trials to learn a problem correlates strongly with the magnitude of change in the synaptic weights, which manifest as changes to the vector field that shape population dynamics. Faster learning on later problems is therefore a result of smaller changes in the synaptic weights that elicit smaller changes in the vector field. A general insight from these results is that changes in the vector field accumulate over the course of learning several problems, effectively decreasing the necessary changes in the vector field to learn a new problem, and thereby accelerating learning. Collectively, these results offer a mechanistic explanation for learning-to-learn by showing that it stems from the formation of latent representations that scaffold learning, and the cumulative changes to population dynamics from early learning that facilitate later learning.

Biosketch

Xiao-Jing Wang is Distinguished Global Professor of Neural Science, director of the Swartz Center for Theoretical Neuroscience at New York University. Previously he was Professor at Brandeis University and Yale University. Dr. Wang’s research focuses on theory and neural mechanisms of cognitive functions such as working memory and decision-making, with a special interest in the prefrontal cortex (often called the “CEO of the brain”). More recently, his group developed biologically-based modeling of large-scale brain circuits. His work bridges with AI in the areas of training recurrent neural networks to perform many cognitive tasks and learning to learn.
James J. DiCarlo WEBSITE

Massachusetts Institute of Technology

Reverse Engineering Visual Intelligence

The brain and cognitive sciences are hard at work on a great scientific quest — to reverse engineer the human mind and its intelligent behavior. Yet these field are still in their infancy. Not surprisingly, forward engineering approaches that aim to emulate human intelligence (HI) in artificial systems (AI) are also still in their infancy. Yet the intelligence and cognitive flexibility apparent in human behavior are an existence proof that machines can be constructed to emulate and work alongside the human mind.

I believe that these challenges of reverse engineering human intelligence will be solved by tightly combining the efforts of brain and cognitive scientists (hypothesis generation and data acquisition), and forward engineering aiming to emulate intelligent behavior (hypothesis instantiation and data prediction). As this approach discovers the correct neural network models, those models will not only encapsulate our understanding of complex brain systems, they will be the basis of next-generation computing and novel brain interfaces for therapeutic and augmentation goals (e.g, brain disorders).

In this session, I will focus on one aspect of human intelligence — visual object categorization and detection — and I will tell the story of how work in brain science, cognitive science and computer science converged to create deep neural networks that can support such tasks. These networks not only reach human performance for many images, but their internal workings are modeled after— and largely explain and predict — the internal workings of the primate visual system. Yet, the primate visual system (HI) still outperforms current generation artificial deep neural networks (AI), and I will show some new clues that the brain and cognitive sciences can offer.

These recent successes and related work suggest that the brain and cognitive sciences community is poised to embrace a powerful new research paradigm. More broadly, our species is the beginning of its most important science quest — the quest to understand human intelligence — and I hope to motivate others to engage that frontier alongside us.

Biosketch

James DiCarlo is a Professor of Neuroscience, and Head of the Department of Brain and Cognitive Sciences at the Massachusetts Institute of Technology. His research goal is to reverse engineer the brain mechanisms that underlie human visual intelligence. He and his collaborators have revealed how population image transformations carried out by a deep stack of neocortical processing stages — called the primate ventral visual stream — are effortlessly able to extract object identity from visual images. His team uses a combination of large-scale neurophysiology, brain imaging, direct neural perturbation methods, and machine learning methods to build and test artificial neural network models of the ventral visual stream and its support of cognition and behavior. Such an engineering-based understanding is likely to lead to new artificial vision and artificial intelligence approaches, new brain-machine interfaces to restore or augment lost senses, and a new foundation to ameliorate disorders of the mind.
Yukiyasu Kamitani WEBSITE

Kyoto University
ATR

神谷之康

京都大学
ATR脳情報研究所

Reconstructing Visual and Subjective Experience from the Brain

The internal visual world is thought to be encoded in hierarchical representations of the brain. However, previous attempts to visualize perceptual contents based on machine-learning analysis of fMRI patterns have been limited to reconstructions with low level image bases or to matchings to exemplars. While categorical decoding of imagery contents has been demonstrated, the reconstruction of internally generated images has been challenging. I introduce our recent study showing that visual cortical activity can be decoded (translated) into the hierarchical features of a pre-trained deep neural network (DNN) for the same input image, providing a way to make use of the information from hierarchical visual features. Next I present a novel image reconstruction method, in which the pixel values of an image are optimized to make its DNN features similar to those decoded from human brain activity at multiple layers. Our method was able to reliably produce reconstructions that resembled the viewed natural images. While our model was solely trained with natural images, it successfully generalized to artificial shapes, indicating that our model was not simply matching to exemplars. The same analysis applied to mental imagery demonstrated rudimentary reconstructions of the subjective content. Our method can effectively combine hierarchical neural representations to reconstruct perceptual and subjective images, providing a new window into the internal contents of the brain.

Biosketch

Yukiyasu Kamitani, Ph.D.. Professor at Kyoto University and Department Head and ATR Fellow at ATR Computational Neuroscience Laboratories. He received B.A. in Cognitive Science from University of Tokyo, and Ph.D. in Computation and Neural Systems from California Institute of Technology. He continued his research in cognitive and computational neuroscience at Harvard Medical School and Princeton University. In 2004 he joined ATR Computational Neuroscience Laboratories, and since 2015 he is Professor at Kyoto University. He is a pioneer in the field of “brain decoding”, which combines neuroimaging and machine learning to translate brain signals to mental contents. He was named Research Leader in Neural Imaging on the 2005 “Scientific American 50”, and received many awards including Tsukahara Memorial Award (2013), JSPS Prize (2014), and Osaka Science Prize (2015). In 2018, he was selected as an ATR Fellow. He is recently engaged in activity in contemporary art with Pierre Huyghe (“UUmwelt” at Serpentine Galleries, London, 2018) and other artists.
Rosalyn Moran WEBSITE

Department of Neuroimaging, IOPPN,King’s College London

The Free Energy Principle and Active Inference in Silico and in Vivo, Visual Sampling and ‘World Model’ Building

The theory of Active Inference proposes that all biological agents retain self-ness by minimizing their long-term average surprisal. In information theoretic terms, Free Energy provides a soluble approximation to this long-term surprise ‘now’ and necessitates the development of a generative model of the environment within the agent itself. The minimization of this quantity via a gradient flow is purported to be the purpose of neuronal activity in the brain and thus provides a mapping from brain activity to their first-principle computations.

In this talk I will outline the theory of Active Inference and describe how discrete and continuous-time systems that perceive and act can be built in silico, while providing evidence for these implementations in neurobiological and behavioral recordings.

Using two experiments in human participants, I aim to demonstrate that human visual search and classification of the MNIST dataset (experiment 1) and world model building and adjustment in a maze task (experiment 2) can be cast as Active Inference processes that utilize neurobiologically plausible architectures comprising prediction in visual hierarchies and alterations in precision via neuromodulation.

From experiment 1, I will show how the expected variational free energy over future states can be used by computer vision systems to decide when and where to look and that these in silico properties are observed in human visual search. By measuring eye-movement responses to the classic Machine Vision dataset – MNIST we will show that humans decide when and where to look, sample data and draw conclusions in a way commensurate with this theory. I will contrast this scheme to popular machine vision systems that use solely feedforward networks and discuss the implications for human-like computing.

From experiment 2, we will provide a flexible model updating regime that carries the hallmarks of the noradrenergic system. By performing tasks including maze search and go/no-go we can demonstrate a specific role for noradrenaline that enables an Active Inference agent to flexibly update its model of current environmental contingencies. We will show emergent features of the in silico system that mimic neuronal firing patterns of noradrenergic cells in the locus coeruleus and behaviour that mimics inter-individual search phenotypes.

Biosketch

Dr. Moran’s research in computational neuroscience combines theoretical and empirical approaches to understand the principles of brain function. In particular, her work is focused on cognition and computational psychiatry – an emerging discipline that aims to use modeling techniques to understand and better treat psychiatric disorders.
Her theoretical and methodological work takes Bayesian approaches to analyse brain dynamics and networks using dynamic causal modeling (DCM). She is a co-author of the academic software package Statistical Parametric Mapping (a standard in brain imaging) and serve on its faculty at courses internationally.
Her lab applies the Free Energy principle as a principle to develop new methods in Artificial Intelligence and in disease modelling. Diseases of interest include age-related neurodegenerative disease and schizophrenia.
Dr Moran trained as an Electrical & Electronic Engineer and obtained her PhD in Engineering from University College Dublin. She trained in neuroscience at University College London under Professor Karl Friston. She has held faculty positions at Virginia Tech and the University of Bristol. She is currently deputy head of the Department of Neuroimaging at the Institute of Psychiatry, Psychology and Neuroscience at King’s College London where she is a Reader of Theoretical Neurobiology.
Terrence Sejnowski WEBSITE

Salk Institute & Univesity of California San Diego

Artificial Intelligence Meets Human Intelligence

AI and neuroscience are converging. This is a consequence of recent developments in deep learning, based on learning algorithms for neural networks from the 1980s. Advances in neuroscience can now inspire new architectures for AI and advances in AI can inform our understanding of brain function. This is illustrated by a Mixture of Experts gating architecture that is a model for how the prefrontal cortex encodes and stores numerous, often disparate, schemas and flexibly switches between them. How the brain creates new schemas while preserving and utilizing old ones is unknown. The recurrent gating model of the prefrontal cortex grows adaptively and can naturally reproduce flexible switching behavior and overcome the catastrophic interference that plagues simpler network models when new experiences are learned. This gating network architecture also exhibits transfer learning and robust memory savings, providing a fundamental framework for how the prefrontal cortex may handle new schemas necessary to navigate in a changing world.

Biosketch

Sejnowski received his Ph.D. in physics from Princeton University.He was a postdoctoral fellow in the Department of Neurobiology at Harvard Medical School before joining the faculty at Johns Hopkins,where he was a Professor of Biophysics. He currently holds the Francis Crick Chair at The Salk Institute for Biological Studies, and is a Distinguished Professor of Biology and Computer Science and Engineering at the University of California, San Diego, where he is co-director of the Institute for Neural Computation. He also is the founding editor-in-chief of Neural Computation. His book on “The Computational Brain” introduced distributed representations in neural networks to a generation of neuroscientists.

Sejnowski is a member of the National Academy of Sciences, the National Academy of Engineering, the National Academy of Medicine and the National Academy of Inventors, one of only three living scientists elected to all four national academies. He has written over 500 scientific papers and 15 books and his h-factor is 156.

Sejnowski pioneered learning algorithms for neural networks in the 1980s, inventing the Boltzmann machine with Geoffrey Hinton; this was the first learning algorithm for multilayer neural networks and laid the foundation for deep learning. He trained a network to pronounce English text, called NETtalk, which was the first real-world application of neural networks.

He developed an unsupervised learning algorithm for blind source separation called Independent Component Analysis (ICA), which has many practical applications. This paper has been cited by other researchers over 10,000 times. Sejnowski is the President of the Neural Information Processing Systems (NeurIPS) Foundation, which organizes an annual conference that was attended by over 14,000 researchers in 2019. This community is responsible for the early research on learning algorithms for neural networks and the recent rise of deep learning as a foundational technology for modern artificial Intelligence. His recent book on “The Deep Learning Revolution” is the first book written by an insider for a general audience.
Hidehiko Takahashi WEBSITE

Tokyo Medical and Dental University

高橋英彦　

東京医科歯科大学

Interface between AI and Psychiatry Research

The diagnostic criteria for psychiatric disorders in the Diagnostic and Statistical Manual of Mental Disorders (DSM) and the International Classification of Disease (ICD) are based on the patient’s behavioral signs and symptoms. Therefore, there is an explanatory gap between phenomenological entities and neurobiological underpinnings. Recently, researchers have begun to bridge this explanatory gap using a dimensional approach. This dimensional approach is compatible with data-driven approaches. Using machine learning methods, we built a model (biomarker) to predict diagnostic labels of psychiatric disorders from resting-state functional connectivity MRI (rs-fcMRI). We developed a schizophrenia biomarker from Japanese cohort, and it was successfully generalized to independent validation cohorts in the USA and Europe. We also built a prediction model of working memory capacity from rs-fcMRI, and succeed in predicting working memory deficit in schizophrenia. Finally, if time allows, we would like to introduce our recent research to visualize altered semantic mapping in the brain of schizophrenia patients, which might be relevant to “loosening of associations”, the fundamental symptoms of schizophrenia.

Biosketch

Professor and Chair of Departments of Psychiatry and Behavioral Sciences Graduate School of Medical and Dental Sciences, Tokyo Medical and Dental University (TMDU). 1997 M.D. from TMDU, 2005 Ph.D. from TMDU. 2006 Chief Researcher National Institute of Radiological Sciences, 2008 California Institute of Technology Visiting Associate, 2010 Associate professor of Department of Psychiatry, Kyoto University. His research field is cognitive neuroscience of emotion, decision-making and their disorders. He received several awards including, Bälz Prize (2009) Japan Society for the Promotion of Science Prize (2014).
Anne Churchland WEBSITE

University of California, Los Angeles

Single-trial Neural Dynamics are Dominated by Richly Varied Movements

When experts are immersed in a task, do their brains prioritize task-related activity? Most efforts to understand neural activ- ity during well-learned tasks focus on cognitive computations and task-related movements. We wondered whether task-per-forming animals explore a broader movement landscape and how this impacts neural activity. We characterized movements using video and other sensors and measured neural activity using widefield and two-photon imaging. Cortex-wide activity was dominated by movements, especially uninstructed movements not required for the task. Some uninstructed movements were aligned to trial events. Accounting for them revealed that neurons with similar trial-averaged activity often reflected utterly different combinations of cognitive and movement variables. Other movements occurred idiosyncratically, accounting for trial-by-trial fluctuations that are often considered ‘noise’. This held true throughout task-learning and for extracellular Neuropixels recordings that included subcortical areas. Our observations argue that animals execute expert decisions while performing richly varied, uninstructed movements that profoundly shape neural activity.

Biosketch

Anne Churchland’s laboratory investigates the neural circuits that support decision-making. Understanding how brains makes decisions is critical for gaining insight into complex cognition. When making decisions, humans and animals can flexibly integrate multiple sources of information before committing to action. The ability to flexibly use incoming information distinguishes decisions from reflexes, offering a tractable entry point into more complex cognitive processes defined by flexibility, such as abstract thinking, reasoning and problem solving. To understand the neural mechanisms that support decision-making, the lab measures and manipulates neurons in cortical and subcortical areas while animals make decisions about sensory signals. To connect the neural responses with behavior, they use mathematical analyses aimed at understanding what information is represented at the level of neural populations, both at a given moment and over time. Understanding neural population activity will bolster the long term goal of understanding how the brain can make decisions that integrate inputs from our multiple senses, stored memories and innate impulses.
Kenji Doya WEBSITE

Okinawa Institute of Science and Technology

銅谷賢治

沖縄科学技術大学院大学

Toward the Society of AI Agents: What Should We Learn from the Brain and Human Society

A major advantage of AI agents compared to humans is that their brains are directly connected to the internet. That makes magics like telepathy, brain transplant and Lamarckian evolution quite easy for AI agents. Then what is the right communication protocols for AI agents, for efficiency and security? Should they stick to human readable languages, or can they evolve own languages? Some of the AI agents can be highly advanced with own curiosity, imagination and goal setting. Such autonomous AI agents may bring new scientific discoveries, invent new technologies, or create new industries, which is fantastic but we have to be concerned about possible dangers.

In assessing and avoiding the dangers in the society of AI agents, learning from the history of human society would be helpful, because humankind is the most dangerous species on this planet. By learning from all kinds of disasters in history, human society has developed a variety of mechanisms to avoid catastrophes from runaway or abuse of intelligent power. For example, rules of democracy, such as election, positions with fixed term, separation of three powers, and local governance, are based on the historical lesson that concentration of unlimited power to certain individuals and groups lead to corruption and repression. Modern economic rules like antitrust law and the right to strike also avoid over-concentration of powers. Peer-reviewing in science has also been the basis for its sustainable progress.

The wisdom gained in the human society can be valuable in preventing runaway and abuse in the society of AI agents. It is important not to let a single AI system to control the whole society. Creating a peer-reviewing mechanism across multiple open-sourced AI agents would be an important norm for establishing reliable and sustainable AI society.

In doing so, the insights form the human brain that created our society should be helpful. For example, the human brain has inherent mechanisms that promote sympathy and avoid inequality, which may constrain advanced intelligence that can support rational, self-interested actions. A genius with no social skills can rarely succeed in real life. Advanced AI agents should also have social competence to be trusted and integrated in the society, and that can be another area where learning from the brain is desired.

Biosketch

Kenji Doya is a Professor at Neural Computation Unit, Okinawa Institute of Science and Technology Graduate University. He took BS in 1984, MS in 1986, and Ph.D. in 1991 at U. Tokyo. He became a research associate at U. Tokyo in 1986, U. C. San Diego in 1991, and Salk Institute in 1993. He joined Advanced Telecommunications Research International (ATR) in 1994 and became the head of Computational Neurobiology Department, ATR Computational Neuroscience Laboratories in 2003. In 2004, he was appointed as the Principal Investigator of Neural Computation Unit, Okinawa Institute of Science and Technology (OIST) and started Okinawa Computational Neuroscience Course (OCNC) as the chief organizer. As OIST established itself as a graduate university in 2011, he became a Professor and served as the Vice Provost for Research till 2014. He serves as the Co-Editor in Chief of Neural Networks since 2008 and a board member of Japanese Neural Network Society (JNNS) and Japan Neuroscience Society (JNSS). He served as the Program Co-Chair of International Conference on Neural Information Processing (ICONIP) in 2007 and 2016, the Program Chair of JNSS meeting in 2010, and the General Chair of JNNS meeting in 2011 and 2018. He received Tsukahara Award and JSPS Award in 2007, MEXT Prize for Science and Technology in 2012, INNS Donald O. Hebb Award in 2018, JNNS Academic Award, APNNS Outstanding Achievement Award, and the age-group 2nd place at Ironman Taiwan in 2019.
Arisa Ema WEBSITE

The University of Tokyo Institute for Future Initiatives

江間有沙

東京大学未来ビジョン研究センター

Interpretative Flexibility’ in AI and Neuroscience Research

The search for artificial intelligence (AI) and brain science that enables “human-like” flexible behavior and communication also raises questions at the interface between science, technology, and society. Assuming we are a collection of information or data, we can build “human-like” AI. As we live, we leave our footprints in the real world and in digital spaces. Some data are biologically assigned, while others are socially assigned. Aspects such as hobbies and personalities may be said to be decentralized information formed between relationships among people. By scraping these pieces of information and data together, zombies are created. However, collecting information does not necessarily make one an individual organism. On the other hand, we should not forget that, as Andy Clark points out, we ourselves are “natural-born cyborgs.” Living in a modern society means living as a cyborg, and becoming a grainy zombie after death, constantly sprinkling the material to create a golem using parts of you. To protect the dignity and value of human beings while advancing technology, multi-stakeholder discussions in various fields such as philosophy, law, and society are required. What kind of society would we like to live in? This is not a question that can be answered easily. Therefore, we must consider responsibility of researchers and continue dialogue with the general public.

Biosketch

Arisa Ema is a Project Assistant Professor at the University of Tokyo and Visiting Researcher at RIKEN Center for Advanced Intelligence Project in Japan. She is a researcher in Science and Technology Studies (STS), and her primary interest is to investigate the benefits and risks of artificial intelligence by organizing an interdisciplinary research group. She is a co-founder of Acceptable Intelligence with Responsibility Study Group (AIR) (http://sig-air.org/) established in 2014, which seeks to address emerging issues and relationships between artificial intelligence and society. She is a member of the Ethics Committee of the Japanese Society for Artificial Intelligence (JSAI), which released the JSAI Ethical Guidelines in 2017. She is also a board member of the Japan Deep Learning Association (JDLA) and chairing Public Affairs Committee. She was also a member of the Council for Social Principles of Human-centric AI, The Cabinet Office, which released “Social Principles of Human- Centric AI” in 2019. She obtained Ph.D. from the University of Tokyo and previously held a position as Assistant Professor at the Hakubi Center for Advanced Research, Kyoto University.
Hiroaki Kitano WEBSITE

Okinawa Institute of Science and Technology

北野宏明

沖縄科学技術大学院大学

Nobel Turing Challenge : A Grand Challenge on AI for Scientific Discovery

One of the most exciting and disruptive research in AI is to develop AI systems that can make major scientific discovery by itself with high-level of autonomy. In this talk, I propose “Nobel Turing Challenge” to the grand challenge bridging AI and other scientific communities. The challenge calls for development of AI systems that can make major scientific discoveries some of which worth Nobel Prize, and the Nobel Committee, and the rest of the scientific community, may not be able to distinguish if it was discovered by human scientist or AI (Kitano, H., AI Magazine, 37(1) 2016). This challenge is particularly significant in biomedical domain where progress of systems biology (Kitano, H., Science, 295, 1662-1664, 2002; Kitano, H., Nature, 420, 206-210, 2002) resulted in total overflow of data and knowledge far beyond human comprehension. After 20 years of my journey in systems biology research, I have concluded that next major breakthrough in systems biology requires AI-driven scientific discovery. Initially, it shall be introduced as AI-assisted science, but it will result in AI scientists with high-level of autonomy. This challenge poses a series of fundamental questions on the nature of scientific discovery, limits of human cognition, implications of individual paths toward major discoveries, computational meaning of serendipity or scientific intuition, and many other issues that may bring AI research into the next stage.

References

Kitano, H. Artificial Intelligence to Win the Nobel Prize and Beyond: Creating the Engine for Scientific Discovery. AI Magazine, 37(1), 39-49, 2016 Kitano, H. Systems Biology: A Brief Overview. Science. 295(5560), 1662-1664, 2002 Kitano, H. Computational Systems Biology. Nature. 420(6912), 206-210, 2002

Biosketch

Dr. Hiroaki Kitano is Professor at Okinawa Institute of Science and Technology Graduate University, President of The Systems Biology Institute, President and CEO of Sony Computer Science Laboratories, Inc., CEO of Sony AI Inc., and Corporate Executive of Sony Corporation. He is also a Founding President of the RoboCup Federation, President of International Joint Conference on Artificial Intelligence (IJCAI) (2009-2011) and Member of the AI & Robotics Council of World Economic Forum (2016-2018). He received The Computers and Thought Award from the International Joint Conference on Artificial Intelligence in 1993, Prix Ars Electronica 2000, Design Award 2001 from Japan Inter-Design Forum, and Nature Award for Creative Mentoring in Science (Mid Carrier) in 2009, as well as being an invited artist for Biennale di Venezia 2000 and Museum of Modern Art, New York in 2001.
Stuart Russell WEBSITE

University of California, Berkeley

1. Brains, Circuits, and Things

I will argue that in a world made largely of things,　representations made of circuits are inadequate. Greater expressive power,　such as that provided by probabilistic programming systems, is required. Integration of this rapidly expanding technology with deep networks is likely to provide the next big step forward in AI.

2. Human Compatible Artificial Intelligence

The question of what happens when AI succeeds in its quest for true machine intelligence is seldom considered. Alan Turing was not optimistic:
”It seems probable that once the machine thinking method had started, it would not take long to outstrip our feeble powers. …At some stage therefore we should have to expect the machines to take control.”

This is the problem of control: if we create machines more powerful than ourselves, how do we retain power over them forever? Conversely, and perhaps more positively, we would like to create provably beneficial AI : AI systems that are guaranteed to be of benefit to humans, no matter how capable they become. This is the essence of trustworthy AI: we trust a machine if and only if we have good reason to believe it will act in ways beneficial to us.

The mistake in the standard model of AI is the assumption that we humans can supply a complete and correct definition of our true preferences to the machine. From the machine’s point of view, this amounts to the assumption that the objective it is pursuing is exactly the right one. We can avoid this problem by defining the goals of AI in a slightly different way:
Machines are beneficial to the extent that their actions can be expected to achieve our objectives.

After all, we don’t want intelligent machines if they are not beneficial, so perhaps we should have pursued this formulation all along. It is more difficult from the machine’s point of view, because the objectives remains within us, not in the machine. It is, nonetheless, feasible. Informally, it translates into three principles:

• The machine’s only objective is to maximize the realization of human preferences.
• The machine is initially uncertain about what those preferences are.
• The ultimate source of information about human preferences is human behavior.
The key characteristics of the new model are the absence of a fixed, known objective—whether at design time or embedded in the agent itself—and the flow of preference information from human to machine at runtime. Here, “preferences” refers to everything that people, in all their variety, might care about with regard to how the future unfolds.

The new model is strictly more general than the standard model, because uncertainty about preferences has certainty as a special case. It is also at least as amenable to instantiation in a wide variety of forms. One particular formal instantiation is the assistance game—originally a cooperative inverse reinforcement learning or CIRL game.

Under reasonable assumptions, the robot will defer to human requests, choose “minimally invasive” strategies that change the world as little as possible, ask permission before taking risky actions whose outcomes may violate human preferences, and, perhaps most importantly, allow itself to be switched off. Even if it knows nothing about human preferences, it can still take actions that improve the human’s freedom of action. All of these behaviors follow from the definition of the problem that the robot is solving, and not from any human-supplied scripts.

Biosketch

Stuart Russell received his B.A. with first-class honours in physics from Oxford University in 1982 and his Ph.D. in computer science from Stanford in 1986. He then joined the faculty of the University of California at Berkeley, where he is Professor (and formerly Chair) of Electrical Engineering and Computer Sciences, holder of the Smith-Zadeh Chair in Engineering, and Director of the Center for Human-Compatible AI. He has served as an Adjunct Professor of Neurological Surgery at UC San Francisco and as Vice-Chair of the World Economic Forum’s Council on AI and Robotics. He is a recipient of the Presidential Young Investigator Award of the National Science Foundation, the IJCAI Computers and Thought Award, the World Technology Award (Policy category), the Mitchell Prize of the American Statistical Association, the Feigenbaum Prize of the Association for the Advancement of Artificial Intelligence, and Outstanding Educator Awards from both ACM and AAAI. From 2012 to 2014 he held the Chaire Blaise Pascal in Paris, and he has been awarded the Andrew Carnegie Fellowship for 2019 to 2021. He is an Honorary Fellow of Wadham College, Oxford; Distinguished Fellow of the Stanford Institute for Human-Centered AI; Associate Fellow of the Royal Institute for International Affairs (Chatham House); and Fellow of the Association for the Advancement of Artificial Intelligence, the Association for Computing Machinery, and the American Association for the Advancement of Science. His book “Artificial Intelligence: A Modern Approach” (with Peter Norvig) is the standard text in AI; it has been translated into 14 languages and is used in over 1400 universities in 128 countries. His research covers a wide range of topics in artificial intelligence including machine learning, probabilistic reasoning, knowledge representation, planning, real-time decision making, multitarget tracking, computer vision, computational physiology, and philosophical foundations. He also works for the United Nations, developing a new global seismic monitoring system for the nuclear-test-ban treaty. His current concerns include the threat of autonomous weapons and the long-term future of artificial intelligence and its relation to humanity. The latter topic is the subject of his new book, “Human Compatible: AI and the Problem of Control” (Viking/Pengun, 2019).