Speakers

Maneesh Sahani WEBSITE

Gatsby Computational Neuroscience Unit

Statistical learning and inference in the perceiving brain

Perception is the process by which we and other animals make sense of the world from sensory input. A hallmark of natural perception is its flexibility. Observers are able to parse sensory input into constituent components at many levels of detail and interrogate their internal representations to report (or act on the basis of) features at any level. Furthermore, they learn to do so without curated datasets and, at least in the case of non-human animals, without access to linguistic training labels. We are still far from understanding the processes that underlie such learning and inference, but some basic principles of statistical learning and inference are very likely to apply. I will review our understanding of how such processes may be instantiated in neural circuits, and point to directions of current progress.

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.
Angela Langdon WEBSITE

National Institute of Mental Health / NIH

[Online]

Richly structured reward predictions in the learning circuits of the brain

Theories from reinforcement learning have been highly influential for interpreting neural activity in the biological circuits critical for animal and human learning. Central among these is the identification of phasic activity in dopamine neurons as a reward prediction error signal that drives learning in basal ganglia and prefrontal circuits. However, recent findings suggest that dopaminergic prediction error signals have access to complex structured reward predictions, and are sensitive to more properties of outcomes than learning theories with simple scalar value predictions might suggest. These findings suggest biological agents learn rich associative structure from their environment and leverage it to build reward predictions that include information about the specific, and perhaps idiosyncratic, features of available outcomes, using these to guide behavior in even quite simple reward learning tasks.

Biosketch

Angela Langdon is an Investigator in the Intramural Research Program of the National Institute of Mental Health at the National Institutes of Health, where she leads the Unit on the Neural Computations in Learning. Research in her group focuses on the development, testing and refinement of biologically plausible algorithms from reinforcement learning that give insight into how animals and humans learn to predict events and outcomes directly from observing and interacting with their environment, how these algorithms are instantiated in the reward learning circuitry of the brain, and how critical changes to this circuitry can cause learning computations to go awry in disorders of mental health.
Xiao-Jing Wang WEBSITE

New York University

[Online]

Deep Recurrent Neural Networks: What can large-scale brain modeling tell us?

In spite of considerable progress in computational neuroscience, most theoretical research has been limited to local neural circuits. By contrast, remarkable advances in AI rely on deep, albeit feedforward, neural networks. With recent technological advances, neuroscience of multi-regional brain-wide neural circuits is taking off. Here I will introduce large-scale modeling of primate cortex based on mesoscopic connectome, which shows that inter-areal connectivity is dense and endowed with an abundance of feedback loops. Our model naturally gives rise to a hierarchy of timescales with numerous functional implications. I will highlight the concept of macroscopic gradients of synaptic excitation and inhibition as a general principle of large-scale cortical organization, and illustrate how our model can be used to investigate distributed cognitive functions such as working memory and decision making. These primarily findings suggest new ideas toward a brain-inspired framework for deep recurrent neural networks in the field of machine learning.

Biosketch

Xiao-Jing Wang is Distinguished Global Professor of Neural Science and director of the Swartz Center for Theoretical Neuroscience at New York University, formally Professor at Yale University. Dr. Wang is known for his pioneering work on the neural mechanisms of working memory and decision-making, multi-regional large-scale brain circuit modeling, and computational psychiatry. Among recognitions he has received are Guggenheim Fellowship, Swartz Prize for Theoretical and Computational Neuroscience Prize, Goldman-Rakic Prize for Outstanding Achievement in Cognitive Neuroscience, election to the Royal Academy of Belgium.
Terrence Sejnowski WEBSITE

Salk Institute Univesity of California San Diego

[Online]

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.
Tetsuya Ogata WEBSITE

Waseda University

Intelligence of Real Robots Based on Predictive Coding Principle

In the machine learning framework represented by deep learning, the final goal is to acquire a model with optimal generalization performance by training big data. However, it is impossible for real robots that perform the tasks in the real world to acquire such optimal policies that can handle all situations. Here, the principle of predictive coding not only trains models to minimize prediction error, but also dynamically adapts perceptions and actions. This talk will introduce our proposed “deep predictive learning” framework. In this framework, multiple prediction models in a real robot consisting of multiple deep learning models represent sensory-motor experiences as multiple attractors, and entrain them with external sensory and motor input to minimize prediction errors. As a result of this entertainment, the model coordinates (controls) its own motion while creating virtual perceptions in real time that differ from actual ones. In addition, the results of robotics research using this deep predictive learning and examples of joint research with several companies will be presented.

Biosketch

Tetsuya Ogata received the B.S., M.S., and D.E. degrees in mechanical engineering from Waseda University, Tokyo, Japan, in 1993, 1995, and 2000, respectively. He was a Research Associate with Waseda University from 1999 to 2001. From 2001 to 2003, he was a Research Scientist with the RIKEN Brain Science Institute, Saitama, Japan. From 2003 to 2012, he was an Associate Professor with the Graduate School of Informatics, Kyoto University, Kyoto, Japan. Since 2012, he has been a Professor with the Faculty of Science and Engineering, Waseda University. From 2017, he is a Joint-appointed Fellow with the Artificial Intelligence Research Center, National Institute of Advanced Industrial Science and Technology, Tokyo.
Tom Macpherson WEBSITE

Osaka University

Neural Circuit Mechanisms in Behavioral Flexibility

The ability to use environmental cues to flexibly guide responses is crucial for adaptive behavior and is thought to be controlled within a series of cortico-basal ganglia-thalamo-cortical loops. At the heart of these loops, the striatum, the major input nucleus of the basal ganglia, is predicted to dynamically guide behavior via signaling through two largely distinct output pathways, commonly referred to as “direct” and “indirect” pathways due to their routes through the basal ganglia. In a series of experiments, we reveal the contributions of these two pathways to reward learning and behavioral flexibility. Specifically, we demonstrate that while activity in “direct” pathway neurons in the nucleus accumbens (NAc), the ventral portion of the striatum, is necessary for some forms of reward learning, activity in “indirect” pathway NAc neurons is required for some, but not all forms, of behavioral flexibility. These studies begin to elucidate the nuanced roles of striatal output pathways in specific forms of learning and demonstrate the complexity of flexible behavioral control in cortico-basal ganglia-thalamo-cortical loops.

Biosketch

Tom Macpherson received his PhD in Behavioral and Clinical Neuroscience from University of Sussex, UK (2009-2014), studying the neural mechanisms of cocaine addiction in mouse models. Following completion of his PhD, he was awarded a JSPS Postdoctoral Fellowship for Research in Japan at Kyoto University (2014-2018). During this time, his research projects led to several major discoveries concerning the roles of Basal Ganglia signaling pathways in reward and aversive learning. Since 2018, Tom has worked as an assistant professor of Behavioral and Clinical Neuroscience at the Institute for Protein Research, Osaka University, using cutting-edge neural imaging, artificial intelligence, genetic, molecular, and behavioral analyses to elucidate the neural mechanisms underlying motor, cognitive, and limbic control within Basal Ganglia neurocircuits.
Jun Izawa WEBSITE

Tsukuba University

Meta-learning for motor learning

Reinforcement learning enables the brain to learn optimal action selection, such as go or no-go, by forming state-action and action-outcome associations. Learning to learn by rewards, i.e., reinforcement meta-learning, is a crucial mechanism for machines to develop flexibility in learning, which is also considered in the brain without empirical examination. Does this mechanism also optimize the brain’s willingness to learn, such as learn or not learn? Meta-learning enables machine learning to regulate hyper-parameters for learning to maximize the long-term goal of learning. Recently, reinforcement learning of learning machines taking advantage of the gradient of the expectation of a long-term goal of learning which exists above the learning machine which cares about a short-term goal of learning has been developed. Scientists also considered such a meta-learning system in the brain. Here, we show that humans learn to learn or not to learn to maximize rewards in visuomotor learning tasks. This regulation of learning is a result of an instrumental, active process that takes into account the learning-outcome structure. Our results thus demonstrate the existence of reinforcement meta-learning in the human brain. Because motor learning is a process of minimizing sensory errors, our findings reveal an essential interaction between reward and error in motor learning.

Biosketch

Jun Izawa received his Ph.D. in Engineering at the Tokyo Institute of Technology, in 2004. After postdoctoral training at Johns Hopkins University, he became a researcher at ATR Computational Neuroscience Laboratories in 2011. Since 2014, he has been an associate professor at the University of Tsukuba. His research interests are in computational motor learning, computational neuro-rehabilitation, and cognition in virtual reality.
Rieko Osu WEBSITE

Waseda University

Hand choice and neuro-rehabilitation

The problem of hand choice in stroke hemiparetic patients is an excellent example of the combination of reinforcement learning and supervised learning. To recover the function, patients must use the paretic hand in their daily life. However, patients prefer to use the non-paretic hand because it gives more rewards, such as quick and accurate reach to food (learned non-use), which will deprive the chance of learning paretic hand function. Therefore, in their subacute rehabilitation process, it is crucial to train to choose the paretic hand, in addition to training its function. In this talk, I will introduce several attempts to understand and intervene the hand choice using behavioral experiments and brain/nerve non-invasive stimulations.

Biosketch

Rieko Osu received her Ph.D. in Psychology at Kyoto University, Faculty of Literature, in 1996. From 1996 to 2001, she was a researcher for ERATO, JST. From 2001 to 2015, she worked as a researcher/senior researcher/department head of ATR Computational Neuroscience Laboratories. From 2015 to 2016, she worked as a director of consumer neuroscience at The Nielsen Company Japan. She has been a professor at Waseda University, Faculty of Human Sciences, since 2017. Her research interests are in computational motor control, neuro-rehabilitation, neuroimaging, and cognitive/affective neuroscience.
Jun Tani WEBSITE

Okinawa Institute of Science and Technology

Cognitive Neurorobotics Studies Using the Frameworks of Predictive Coding and Active Inference

My main research motivation has been to illuminate essential mechanisms for embodied cognition, focusing on its top-down and bottom-up interactions processes. For this purpose, we developed a variational RNN model using the free energy principle, and applied it to a set of robotics experiments, while extending the frameworks of predictive coding and active inference. Our experimental results revealed an interesting finding that the estimated precision develops differently, depending on a parameter called the meta-prior, w , which represents the strength for regulating with which the divergence between the approximate posterior and the prior probability distribution in the free energy formula is regulated. A robot trained with larger w tends to exhibit stronger top-down intention, due to high developed prior precision developed; therefore, it tries to minimize the gap between the top-down and bottom-up processes by changing the external worlds through actions. On the other hand, a robot trained with smaller w tends to change its own intention to minimize the sensory prediction error, due to the low prior precision. By consolidating these findings, I will discuss how the developed precision structure could affect embodied cognitive processes in social interactions, as well as in interactions with physical objects.

Biosketch

Jun Tani received the D.Eng. degree from Sophia University, Tokyo in 1995. He started his research career with Sony Computer Science Lab. in 1993. He became a Team Leader of the Laboratory for Behavior and Dynamic Cognition, RIKEN Brain Science Institute, Saitama, Japan in 2001. He became a Full Professor with the Electrical Engineering Department, Korea Advanced Institute of Science and Technology, Daejeon, South Korea in 2012. He is currently a Full Professor with the Okinawa Institute of Science and Technology, Okinawa, Japan. He is also a visiting professor of The Technical University of Munich. His current research interests include cognitive neuroscience, developmental psychology, phenomenology, complex adaptive systems, and robotics. He is an author of “Exploring Robotic Minds: Actions, Symbols, and Consciousness as Self-Organizing Dynamic Phenomena.” published from Oxford Univ. Press in 2016.
Masashi Sugiyama WEBSITE

RIKEN Center for Advanced Intelligence Project
The University of Tokyo

[Online]

Recent Advances in Classification from Noisy Labels

Supervised learning from noisy output data is one of the classical problems in machine learning. While this is a straightforward task in regression, classification from noisy labels is actually still challenging since noisy big data does not always help. It has been shown that, when the noise transition matrix, which specifies the flipping probability of labels, is available, the bias caused by label noise can be canceled by appropriately correcting the loss function. However, when the noise transition matrix is unknown, its estimation from only noisy labels is not straightforward due to its non-identifiability. In this talk, I will give an overview of recent advances in classification from noisy labels, including joint estimation of the noise transition matrix and a classifier, analysis of identifiability conditions, and extension to instance-dependent noise.

Biosketch

Masashi Sugiyama received a Ph.D. degree in Computer Science from Tokyo Institute of Technology, Japan, in 2001. After experiencing assistant and associate professors at the same institute, he became a professor at the University of Tokyo in 2014. Since 2016, he has concurrently served as Director of the RIKEN Center for Advanced Intelligence Project. His research interests include theories and algorithms of machine learning. He was a recipient of the Japan Academy Medal in 2017 and the Commendation for Science and Technology by the Minister of Education, Culture, Sports, Science and Technology in Japan in 2022.
Yutaka Matsuo WEBSITE

The University of Tokyo

Two-story model of intelligence

The field of deep learning is progressing rapidly, and the importance and effectiveness of self-supervised learning is well recognized. In this talk, I will propose a model of intelligence based on self-supervised learning. First, a world model is created by self-supervised learning through observation and action. The world model is a deep generative model of the external world based on experience, and is useful for planning actions. The proposed model of intelligence is a two-story model, where the world model is the first floor and the second floor is responsible for linguistic input and output. The world model on the first floor is used from the outside, and language drives the world model based on verbal input. In other words, “imagining” is the primitive way of understanding meaning. Since linguistic input and output is a series of symbols, it can be regarded as a Turing machine. For a given linguistic task, a corresponding algorithm is learned. As a function of these learned algorithms, various intelligent processes such as memory, recall, reasoning, and symbol processing are realized. Historically speaking, it is the progress of human civilization that has led to the accumulation of various useful modules through the advancement of language tasks. In current deep learning, there is a fundamental technical problem, which is why these mechanisms are only partially realized. The correspondence between these models of intelligence and the brain will also be discussed.

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. At Japan Society for Artificial Intelligence (JSAI), he has served as Editor-in-chief and the chair of the ELSI committee, and has been a board member since 2020. He is the president of Japan Deep Learning Association (JDLA), and a member of the board of directors at SoftBank Group Corp.
He is working on artificial intelligence, especially on deep learning and web mining.
Aida Nematzadeh WEBSITE

DeepMind

Towards Better Pretrained Representations

In the last few years, large pretrained models have achieved exceptional performance on a range of tasks (without task-specific training) and in various domains (such as language, vision, and multimodal). In this talk, I will argue for a better evaluation paradigm for pretrained representations to better understand the extent to which these representations reflect certain aspects of human cognition. In particular, I will focus on verb understanding, common-sense reasoning, and answering questions about images. I will outline some of the strengths and weaknesses of current pretrained language and multimodal models and discuss directions for future work.

Biosketch

Aida Nematzadeh is a staff research scientist at DeepMind. Her research interests are in the intersection of computational linguistics, cognitive science, and machine learning. Her recent work has focused on multimodal learning and evaluation and analysis of neural representations. Before joining DeepMind, she was a postdoctoral researcher at UC Berkeley. She received a PhD and an MSc in Computer Science from the University of Toronto.
Ryota Kanai WEBSITE

ARAYA Inc.

Consciousness and latent space in the brain and machine

I will present our current research exploring potential connections between consciousness and intelligence. In this paper we examined how to combine theories of consciousness with modern deep learning techniques to translate high-level concepts from consciousness research into more concrete computational concepts. Among the theories we considered were Global Workspace Theory (GWT), Information Generation Theory (ITG), and Attention Schema Theory (AST). From this effort, one of the key conclusions is that consciousness has evolved into a platform for general-purpose intelligence. In this talk, I will present a reinterpretation of the GWT as a shared latent space among multimodal specialist modules and outline a roadmap for implementing a Global Latent Workspace (VanRullen & Kanai, 2021) using deep learning techniques, such as an unsupervised translation of representations across latent spaces. With our re-formulation of GWT, we will discuss the functional merits of having a global workspace as well as its implications for neuroscientific research into the brain and the development of brain-to-brain communication technologies.

Biosketch

Ryota Kanai is the founder & CEO of Araya, Inc. After graduating from the Faculty of Science at Kyoto University in 2000, he received his PhD (Cum Laude) in 2005 from Utrecht University in the Netherlands, where he studied human visual information processing mechanisms. After working as a researcher at California Institute of Technology in the U.S. and University College London in the U.K., and as a JST PRESTO researcher and Associate Professor of Cognitive Neuroscience at the University of Sussex in the U.K., he founded Araya, Inc. and worked full time there since 2015. He is engaged in research on the principles of consciousness in the brain and the implementation of consciousness in AI through the fusion of neuroscience and information theory. He has been also working on the practical application of AI and neuroscience in industry. He has received many awards, including the Young Scientist Award from the Ministry of Education, Culture, Sports, Science and Technology, the JEITA Venture Award (2020), the ET/IoT Technology Award (2019) among others as Araya Inc. From 2020, he is working on the practical application of brain-machine interface as a project manager of the Moonshot Project in the Cabinet Office.
Yoshua Bengio WEBSITE

University Montreal

[Online]

Inspiration from brains for system 2 deep learning

The state-of-the-art in AI is based on deep learning but a clear gap can be found with several important aspects of human intelligence, for the most part related to the challenge of out-of-distribution generalization, the ability to learn quickly a new task and the ability to reason about learned concepts and causality. Many of these aspects may be traced to system 2 cognitive abilities, i.e., to conscious processing involving attention, imagination, working memory and long-term memory. We attempt to translate this gap in terms of inductive biases that benefits brains but that are yet to be fully implemented in deep learning. This machine learning perspective may also offer hypotheses to explain the evolutionary advantages of these elements, for example why the bottleneck constraint of the global workspace theory may actually be beneficial for systematic generalization. We summarize recent work on deep learning architectures motivated by this analysis and propose a novel framework called generative flow networks or GFlowNets to implement several of these inductive biases and a powerful machinery for learned probabilistic inference.

Biosketch

Yoshua Bengio is Full Professor in the Department of Computer Science and Operations Research at Université de Montreal, as well as the Founder and Scientific Director of Mila and the Scientific Director of IVADO. Considered one of the world’s leaders in artificial intelligence and deep learning, he is the recipient of the 2018 A.M. Turing Award with Geoff Hinton and Yann LeCun, known as the Nobel prize of computing. He is a Fellow of both the Royal Society of London and Canada, an Officer of the Order of Canada, and a Canada CIFAR AI Chair. In 2022, he was appointed Knight of the Legion of Honor by France and named co-laureate of Spain’s Princess of Asturias Award for technical and scientific research.
Ai Koizumi WEBSITE

Sony Computer Science Laboratories. Inc.

Anxiety in Brain meets AI for Society

Our brain persistently remembers threatful experiences to prepare us for similar threats in the future. However, in modernized safe societies where similar threats are unlikely to recur, our neural predisposition to cling to threatful memories could be maladaptive and may lead to exaggerated anxiety. How to alleviate preserved anxiety remains one central challenge in neuroscience. Here, I present our neurofeedback projects where we leverage machine learning techniques to modify the neural representations of traumatic memories. Specifically, we used machine learning techniques to reinforce the activation patterns in human visual areas representing threatening objects while bypassing conscious exposure. This neurofeedback technique has successfully reduced fear-like responses with experimentally-induced traumatic memories, subclinical animal phobias, and clinical symptoms among a small set of patients with post-traumatic stress disorders. Those projects demonstrate a case where neuroscience and AI could synergistically augment the means to handle mental health issues. Our additional preliminary projects suggest that neuroscience may further extend the approaches to coping with maladaptive forms of anxiety by perceiving anxiety as an emerging property not only from the brain but also from the individual body and the collective societies.

Biosketch

Ai KOIZUMI received her B.A. in Psychology and Cinema (double major) from The State University of New York, Binghamton (2005), and her Ph.D. in Psychology from the University of Tokyo (2012). Her research career includes time as a researcher in the Columbia University Department of Psychology and the Center for Information and Neural Networks (CiNet). She joined Sony Computer Science Laboratories in January 2019 as an associate researcher. She currently serves as a cooperating researcher with Advanced Telecommunications Research Institute International (ATR) and Precursory Research for Embryonic Science and Technology (PRESTO). Previously, she served as a special researcher or special international researcher with the Japan Society for the Promotion of Science, NTT Communications Science Laboratories, the Japan-U.S. Brain Research Cooperation Program, and the Leading Initiative for Excellent Young Researchers (LEADER).
Patricia Churchland WEBSITE

Salk Institute University of California San Diego

[Online]

Endothermy, Neocortex, & Social Caring in Mammals and Birds

Social neuroscience, especially in the last decade, has made impressive progress in exploring the neural mechanisms regulating social behavior, including consolation behavior, attachment and bonding, aggression, willingness to punish, and the effects of nurturing and social stress on the developing brain. In parallel, behavioral research on nonhuman mammals and birds has revealed the existence of prosocial choice, consolation behavior as well as altruistic behavior. In combination, the research raises the wider question of what these various results signify for our understanding of human social motivation in general and moral motivation in particular. Although moral philosophers have discussed norms and values since Socrates and Confucius, the scientific approach has provided new insights and provoked a reconsideration of common assumptions about the nature and origin of moral values. This talk has five parts: (1) the evolutionary origin of sociality in mammals and birds, (2) a brief geography of moral philosophy (3) presentation of selected highlights from social neuroscience, (4) discussion of the links between the reward system and reinforcement learning of social norm and (5) what this all means for understanding moral values.

Biosketch

Patricia Smith Churchland is a Professor emerita of Philosophy at the University of California, San Diego, and an adjunct Professor at the Salk Institute. Her research focuses on how discoveries in neuroscience impact traditional philosophical ideas. She is author of the pioneering book, Neurophilosophy (MIT Press 1986), and co-author with T. J. Sejnowski of The Computational Brain (MIT 1992). Her current work focuses on morality and the social brain; Braintrust: What Neuroscience tells us about Morality (2011 Princeton U P). Touching a Nerve, published by Norton in 2013, portrays how to get comfortable with this fact: I am what I am because my brain is as it is. In June 2019, her latest book was released – Conscience: The Origins of Moral Intuition. She has been president of the American Philosophical Association and the Society for Philosophy and Psychology, and won a MacArthur Prize in 1991, the Rossi Prize for neuroscience in 2008, and the Prose Prize for science for the book, Braintrust. She was chair of the Philosophy Department at the University of California San Diego from 2000-2007.
Hideaki Shimazaki WEBSITE

Hokkaido University

Consciousness and the thermodynamics of the Bayesian brain

What aspects of neural activity underpin our conscious experiences? While we may need various measures to characterize them, here we propose that seeing the brain as ‘an information-theoretic engine’ might carve out an essential aspect of the consciousness experiences.

Neurophysiological studies on early visual cortices revealed that an initial feedforward-sweep of neural response conveys stimulus features only, whereas modulation of the late component (e.g., ~100 ms after the stimulus onset) presumably mediated by feedback connections from higher brain regions carries perceptual effects such as awareness and attention. Psychophysical experiments on humans using visual masking or transcranial magnetic stimulation showed that selective disruption of the late component vanishes conscious experiences of the stimulus. Here we provide a unified computational and statistical view of the modulation of sensory representation by internal dynamics in the brain.

A key computation is the nonlinear integration of multiple signals known as the gain modulation ubiquitously found in nervous systems as a mechanism to adapt neurons’ nonlinear response functions to stimulus distributions. We show that the Bayesian view of the brain provides a statistical paradigm for gain modulation. More precisely, we can model the delayed gain-modulation of the stimulus-response via recurrent feedback connections as dynamics of the Bayesian inference that combines the observation and top-down prior with time delay. Interestingly, this process becomes a mathematical analog of a heat engine in thermodynamics [1,2]. This view, ‘the brain as an information-theoretic engine,’ allows us to quantify the amount of delayed gain modulation in terms of entropy changes in neural activity, which quantifies the perceptual capacity of neural dynamics. We present a data-driven approach to quantify the delayed gain modulation from spike data using the state-space model of neural populations [3,4].

With the theoretical and methodological advances, we share the view that the thermodynamic approaches to neural systems [4-6] will be critical to uncovering neural activity underlying conscious experiences.
[1] Shimazaki (2015) Neurons as an Information-theoretic Engine. arXiv:1512.07855 (published as a book chapter)
[2] Shimazaki (2020) arXiv:2006.13158
[3] Shimazaki, Amari, Brown, Gruen (2012) PLoS Comp Biol 8(3): e1002385
[4] Donner, Obermeyer, Shimazaki (2017) PLoS Comp Biol 13(1): e1005309
[5] Aguilera, Moosavi, Shimazaki (2021) Nat Commun 12, 1197
[6] Aguilera, Igarashi, Shimazaki (2022) arXiv:2205.09886

Biosketch

Hideaki Shimazaki is a specially-appointed associate professor at the Center for Human Nature, Artificial Intelligence, and Neuroscience (CHAIN) of Hokkaido University in Japan. Before that, he took joint positions in industry and academia as a senior scientist at Honda Research Institute Japan Co., Ltd. and an associate professor at the Graduate School of Informatics in Kyoto University. His research focuses on uncovering the intelligence of organisms and machines using physics, statistics, and machine learning methods, emphasizing data analysis on spiking neural activities. He received M.A. in neuroscience at Johns Hopkins University in 2003 and Ph.D. in physics at Kyoto University in 2007. After his Ph.D., he pursued postdoctoral training in computational neuroscience at RIKEN Brain Science Institute and MIT Department of Brain and Cognitive Sciences.
Keisuke Suzuki WEBSITE

Hokkaido University

Computational Neurophenomenology

Recent developments in deep learning models have significantly scaled up the output of computational models to a quality that we can directly experience with e.g. XR technology. Here, I present Hallucination Machine, a novel virtual reality platform for people to experience synthesised visual hallucinations generated by a deep neural network. We reported marked similarlity between simulated hallucinatory phenomenlogy in Hallucination Machine and psychedelic experiences under psilocybin administration. Based on the results of ongoing surveys with psychedelics users and clinical patients having visual hallucinations, I will discuss a possible neurocomputational mechanism underlying the distinctive perceputual characteristics specifically induced with psychedelic substances. Computational models have been used to explain objective performance of perception and cognition. In contrast, in our approach, which we call ‘computational neurophenomenology’, computational models are emproyed to simulate conscious phenomenology. This appraoch to naturalise phenomenology with new technologies will open a way for computational understanding of conscious experiences.

Biosketch

Keisuke Suzuki obtained his Ph.D degree on the subject of artificial life from the University of Tokyo in 2007. He stayed as a research fellow in RIKEN Brain Science Institute, where he carried out research into human cognitive functions in virtual reality environments (2008-2011). He developed a novel virtual reality system called SR (Substitutional Reality) with his colleagues. In this system, people believe that they are experiencing real-world scenes, even though they are actually experiencing pre-recorded scenes. After leaving Japan, he worked as a research fellow at the Sackler Centre for Consciousness Science in University of Sussex, UK (2011-2021), where he developed various virtual reality platforms for the experimental study of embodied self-consciousness. In 2021, he joined the Center for Human Nature, Artificial Intelligence, and Neuroscience (CHAIN) in Hokkaido University, Japan as a specially appointed lecturer to continue his work in embodied cognition and conscious presence.
One of Keisuke’s key research focuses is about conscious presence; i.e. the subjective feeling of being “here and now”. The sense of presence is one of the important aspects of our subjective conscious experience, but its underlying neural mechanisms remain poorly understood. His approach builds on state-of-the-art virtual reality to experimentally manipulate the bodily and mental states of mind, which is complimented by theoretical modeling.
Makiko Yamada WEBSITE

National Institutes for Quantum Science and Technology

Neural bases of color qualia structure

Most of the major neuroimaging studies on consciousness have used optical illusions. For example, in an ambiguous figure such as the Rubin vase or in a binocular rivalry, only one of the two visual images rises to consciousness and switches at regular intervals. By using such optical illusions, brain activity that changes in correlation with the content of consciousness has been identified even though the physical stimulus is constant. However, in most conventional studies, the quality of consciousness is reduced to a binary problem such as whether or not a specific visual image was seen, the brain activity of qualia itself (for example, the subjective feeling of redness of apples) could not be captured. With the goal of measuring qualia itself, our transformative research areas B “qualia structure” has been working on ways to understand qualia in terms of its relationship to its surroundings. In my talk, I will introduce our basic idea to identify the brain regions corresponding to the qualia structure from the similarity of the brain activity pattern, by combining both the similarity judgment analysis for calculating the similarity of color qualia and the representational similarity analysis for calculating the similarity of brain activity patterns.

Biosketch

Makiko Yamada is a group leader at National Institutes for Quantum Science and Technology (QST), Japan. She is working to elucidate the mechanisms of cognitive biases (cognitive distortions) in the brain that are involved in the development of psychiatric symptoms such as depression and schizophrenia. She received a doctor’s degree in Human Environmental Sciences from Kyoto University in 2006, and joined the Department of Psychiatry, Kyoto University and the Department of Psychology, the University of Chicago as a JSPS PD. In 2009, she began her career in cognitive neuroscience research integrating fMRI and PET at National Institute of Radiological Sciences (NIRS), the predecessor to QST. She is also a visiting associate professor at Tohoku University Graduate School of Medicine, and the Program Manager for the Moonshot Research and Development Program Goal 9.
Tadahiro Taniguchi WEBSITE

Ritsumeikan University

Consciousness and Symbol Emergence Systems

Language is a crucial cognitive capability of humans not only for communication but also for perception and planning. Language is also considered to perform an important role in relation to consciousness. Importantly, symbol systems, including language, are not static but dynamically changing systems. How such symbol systems can be organized, shared, and become used is a long-standing scientific question in language evolution. A major difficulty is that agents cannot inspect each others’ brain states nor transmit meanings directly. There is no feedback about the meaning of a word. Constructive studies of emergent communication based on language games have been studied. Recently, we proposed a Metropolis-Hastings (MH) naming game that achieves emergent communication. The MH naming game models an emergent communication as a decentralized inference procedure of a probabilistic generative model (PGM), which integrates several agents involved in the communication. In this talk, I will talk about the PGM-based approach to emergent communication, i.e., symbol emergence, and the potential relationship between consciousness and the symbol emergence system.

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.
Misako Komatsu WEBSITE

Tokyo Institute of Technology

Cortical information dynamics for predictive coding

Our brain processes the incoming sensory inputs incessantly, and make predictions and generalizations of the environments to adapt our behavior. This process is often referred to as “predictive coding”. To investigate cortical-wide information processing for predictive coding, we developed a whole-cortical electrocorticography (ECoG) array for the common marmoset, a small non-human primate. The array provides an opportunity to capture global cortical information processing with high resolutions at a sub-millisecond order in time and millimeter order in space. In recent work, we have also developed optogenetics system to induce abnormal frontotemporal circuit, which is reported to be associated with schizophrenia. In this talk, I will present (i) an overview of cortical information dynamics for auditory predictive coding process, and (ii) the effects of top-down circuit manipulations on the process.

Biosketch

Misako Komatsu is a Specially Appointed Associate Professor at Tokyo Institute of Technology and a Visiting Researcher at RIKEN Center for Brain Science. She has developed large-scale ECoG, and her current research focus is to understand principles of whole-brain information processing. She received the Ph.D. degree from Tokyo Institute of Technology in 2014. She started her research career at Jun Tani Lab., RIKEN Brain Science Institute as a Research Associate in 2008, and was engaged in research on estimation of network structure of multi-dimensional neural recordings with a Recurrent Neural Network model. Since 2012, she is conducting primate electrophysiology at RIKEN as a Research Associate or Researcher.
Karl Friston WEBSITE

University College London

[Online]

Synaptopathy and the Bayesian brain

This talk considers formal or computational approaches to psychopathology. I will use schizophrenia to offer a case study of computational psychiatry. We first review the basic phenomenology and pathophysiological theories of schizophrenia, with a special focus on synaptopathy and functional dysconnections; i.e., network disorders. These motivate the choice of a formal or computational framework within which to understand psychopathology; particularly, in terms of false beliefs or inference. This framework is the Bayesian brain. We will focus on the (neuromodulatory) encoding of uncertainty or precision within predictive coding implementations of active inference – to demonstrate computational approaches to pathogenesis in neuropsychiatric disorders. The endpoint of this analysis is the key role of neuromodulation in selecting those aspects of the sensorium that underwrite our belief updating – and making sense of our sensations. This speaks to (i) the key role of modulatory neurotransmitters (and psychopharmacological drugs) in shaping our inferences about the lived world, and (ii) non-invasive measurements of context-sensitive connectivity or synaptic efficacy (e.g., with dynamic causal modelling).

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, Collège de France. 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 universities of York, Zurich, Liège and Radboud University.
Yuichi Yamashita WEBSITE

National Center of Neurology and Psychiatry

Psychiatric and neurodevelopmental disorders as failures in hierarchical predictive process: neurorobotics approach

The idea that adaptive behavior and cognition of human are enabled by hierarchical predictive process under generative models (i.e. predictive processing) has become a popular perspective to explain computational principle of the brain. However, while hierarchical predictive process provides significant advantages for adaptive behavior in social environments, their failure to properly develop or maintain precisely aligned signaling in neural systems has been postulated to result in psychiatric or developmental disorder symptoms. In this talk, I will introduce our series of experiments using “neurorobotics approach” in which behavioral control mechanisms with hierarchical predictive process were implemented by the physical actions of a humanoid robot driven by a hierarchical recurrent neural network. These experimental results demonstrated that failures in hierarchical predictive process including functional disconnections between levels of the hierarchical network and altered intrinsic neuronal excitabilities lead to aberrant neural activities, abnormal behavior, and altered estimation of sensory precision that can explain psychiatric and developmental disorders symptoms. We suggest that our methodology of a neural network model-driven robot could provide us new insights for examining the hypothesis of pathological mechanisms in psychiatric and developmental disorders.

Biosketch

Yuichi Yamashita is a Section Chief at Computational Psychiatry Section, Department of Information Medicine, National Institute of Neuroscience, National Center of Neurology and Psychiatry (NCNP). He received M.D. (1998) and Ph.D. (2013) from Tohoku University. After working as a psychiatrist at Department of Neuropsychiatry, Tohoku University Hospital, Department of Psychiatry, and Tokyo Metropolitan Matsuzawa Hospital, he then became a Research Associate at RIKEN Brain Science Institute (Jun Tani’s Lab.) in 2005. Since 2013, he is working with NCNP.
Matthew Botvinick WEBSITE

DeepMind

[Online]

A unifying normative model of dual-process control

Dual-process theories play a central role in both psychology and neuroscience, figuring prominently in fields ranging from executive control to reward-based learning to judgment and decision making. In each of these domains, two mechanisms appear to operate concurrently, one relatively high in computational complexity, the other relatively simple. Why is neural information processing organized in this way? A compelling answer to this question can be derived from the notion of compression. Specifically, dual-process structure emerges naturally in a system that seeks to minimize the description length of its own behavior. In recent work we have applied a single model, based on this idea, to findings from research on executive control, reward-based learning, and judgment and decision making, showing that seemingly diverse dual-process phenomena can be understood as domain-specific consequences of a single underlying set of computational principles.

Biosketch

Director of Neuroscience Research, a team lead in AGI Research at 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.