Seminars

Gašper Tkačik explores whether the language of information in biology can become a predictive scientific theory. The lecture connects information transfer from DNA to proteins, positional signals that guide cell fate during development, neural information processing, and the storage and inheritance of information in evolving genomes. It brings physics, information theory and quantitative biology together to examine these processes across biological scales. Tkačik is Professor at the Institute of Science and Technology Austria. Shinya Kuroda provides commentary; Arisa Ema moderates. Online via Zoom Webinar. Wednesday 14 October 2026, 14:00–15:00 JST (Asia/Tokyo; UTC+9). Public advance registration is required through the organizer’s event page. The lecture is in English with Japanese interpretation. Organized by Tokyo College, The University of Tokyo Institutes for Advanced Study.

Deep networks are often thought of as black boxes. Their ability to encompass vast swathes of knowledge indeed makes them hard to explain. Yet many of their behaviours — generalization under overparametrization, grokking, OOD failures, neural scaling laws — recur across architectures and scales, and each, however surprising, can be reproduced and explained in controlled settings. I will review these efforts to identify and explain the universal phenomena of deep learning, and suggest that an LLM may amount to a sum of such tractable sub-phenomena, interpolative in nature. Finally, leaving scientific rigor aside, I'll argue that what separates this prosaic picture from the apparent magic of LLMs may well be the industrial scale of compute and human labour behind it, and that AGI in its deeper extrapolative sense may be much further away than claimed.

Jie Yang presents BRIDGE, a multilingual evaluation benchmark built from real clinical tasks and more than one million samples derived from electronic health records. The benchmark addresses the gap between simplified examination-style tests and the complexity of clinical data, while enabling comparison of rapidly changing language models. Evaluation of 95 models involved over 24,000 experiments and 39 million predictions. Results vary substantially by model family, task and language, with some open-source systems matching proprietary models. The study also finds that chain-of-thought prompting frequently reduces accuracy on these tasks and examines stigmatizing language produced during model reasoning at large scale.

Matthias Vigl examines whether the computational scaling strategies behind modern machine learning can benefit particle physics. Empirical scaling laws relate model performance to computing resources and help balance model size against training-data volume. For a fixed dataset, increasing model capacity beyond the interpolation threshold can improve generalization through double descent, but gains eventually encounter limits imposed by the data. The talk then considers scaling data and model size together, deriving compute-optimal relations for transformer-based jet taggers trained on as many as billions of simulated jets. It studies how training hyperparameters and input representations change these relations. The observed behaviour suggests that larger computing budgets could yield useful gains in particle physics, especially with a move toward more general-purpose foundation models.

How do humans stabilize an inverted pendulum, and why does a balanced pole eventually fall? Drawing on 25 years of fingertip pole-balancing research, this talk examines neural correction delays: longer poles move more slowly relative to the nervous system’s response time. Delay-differential models can stabilize the upright position, yet skilled people still experience falls. The proposed explanation is microchaos arising from interactions among delay, sensory dead zones and frequency-dependent force encoding. A region of transient falling solutions lies next to stable microchaotic dynamics. Such microchaos is absent in virtual frontal-plane balancing tasks, while models of standing postural sway lack the corresponding transient regime. The comparison suggests that human falls, unlike pole falls, are more plausibly associated with medical events or slips and trips.

To orchestrate complex behaviors in mammals, the cerebral cortex must continuously transmit its processing results to subcortical regions. While internal cortical processing relies on highly selective sparse codes, these descending cortical output streams employ an enigmatic dense code, characterized by high firing rates and low feature selectivity. This dense population code presents a profound paradox: how can downstream regions extract precise, cognitively relevant signals when most cortical output neurons can be active at any moment? We address this paradox by testing the hypothesis that cortex does not lose information when switching to a dense code. Instead, it utilizes a mechanism for sparse-to-dense coding transformations discovered by my laboratory: thalamocortical synapses target specifically the dendritic initiation zone for calcium action potentials (APs), which enables cortical output neurons to transmit multiple information streams simultaneously via a multiplexed 1-2-3 AP syntax. I will provide first evidence that this remarkable synaptic specificity and coding syntax generalize across long-range pathways. Information coupling via an anatomically and biophysically distinct dendritic nexus may hence be a ubiquitous mechanism for sparse-to-dense coding transformations in cortex. Dissecting these mechanistic origins of cortical output streams is only now possible due to a unique in vivo – in silico approach, perfected over two decades, that my laboratory developed for bridging the gaps between dendritic and population-level computations

Yixuan He (Arizona State University) considers graph neural networks for social, financial and biological networks in which edge signs and directions contain information needed for clustering and prediction. After an introduction to graph neural networks, the talk explains how signed and directed methods preserve relationships lost when every connection is treated as unsigned and undirected. The seminar then asks when a representation learned before the final task can transfer to a different graph or to a setting with few labels. Two theoretical perspectives address transfer across graph domains and the use of link prediction for linear community detection. TopoDIG and TopoSIGN combine encoders that preserve edge structure with persistent topological features, enabling pre-training and prompt-based adaptation on directed and signed networks.

Alok Puranik (Jane Street) examines data mixing for large language models: allocating model capacity and computation among training sources. A small model’s preferred data mix may differ from that of a larger model, yet the cost of large-model training forces many decisions to depend on smaller experiments. Scaling laws can predict performance as size and computation increase only when trends remain stable or change predictably. The talk uses experiments from Puranik’s work and frontier laboratories to examine data-mix questions that violate these assumptions, explain why small-scale results can mislead, and explore better methods for choosing training data. His research at Jane Street concerns scaling laws for sequence models in trading.

Navid Azizan presents methods that make neural models obey physical, safety and operational constraints at deployment. Hard-constrained neural networks, or HardNets, enforce input-dependent constraints by construction while preserving universal approximation within the feasible function class. Applications include models of chaotic dynamics with bounded trajectories, energy-constrained operator learning, safe reinforcement learning and control with formal guarantees. The talk then turns to enforcing constraints during sampling from pretrained diffusion and flow-matching models. Formulating generation as trajectory optimization allows receding-horizon control to guide outputs toward feasibility without retraining or imposing excessive restrictions on the sampling process. Examples from fluid dynamics, robot planning and control, PDE control and language-guided image editing illustrate how constraints can define admissible behavior while retaining expressive learning and generation.

Lenka Zdeborová introduces the Rules-and-Facts model, in which some observations follow a shared rule while others are isolated exceptions requiring memorization. The framework studies when a learner can acquire the rule and retain those exceptions simultaneously. Its results emphasize how capacity is organized and deployed, rather than capacity alone. The talk examines how regularization and the geometry of kernels or learned feature maps can reserve resources for memorization without undermining rule learning. It connects the balance between abstraction and memory to architectural and algorithmic choices, providing a theoretical account of neural networks that both generalize and recall specific facts.

MIT Biology Colloquium Series talk by Peter Koo (Cold Spring Harbor Laboratory), hosted by Yunha Hwang, on using AI to learn the gene regulatory code.

Mark Harnett presents work on how synaptic organization, nonlinear dendritic processing, and neuronal activity patterns interact to support computation, flexibility, and learning in the adult mammalian neocortex. The Brain and Cognitive Sciences colloquium is followed by a reception.

Rani Borbara presents brain-wide recordings from more than 150 freely moving C. elegans to examine how anatomical connectivity becomes functional organization during behavior. Cross-animal models map cell-class conditional dependencies and test how synapses, gap junctions, neuromodulation and shared brain states shape circuit dynamics.

Stanford Neurosciences Seminar Series talk by Emery Brown (Massachusetts Institute of Technology, Neuroscience Statistics Research Lab) on deciphering the dynamics of the unconscious brain under general anesthesia, hosted by the Wu Tsai Neurosciences Institute.

Daphna Shohamy and Ashok Litwin-Kumar bridge experimental and computational neuroscience to explain how the brain assigns significance to memories, how dopamine shapes memory-guided decisions, and how bodily signals influence learning. Isabel Low moderates a public discussion and question session.

Zohar Ringel of the Hebrew University of Jerusalem presents a renormalisation-group approach to deep learning based on neural-network field theories. The seminar examines scaling laws, the removal of non-learnable field modes, benign overfitting and open questions about feature learning. The confirmed Quantum Matter seminar takes place in the Bob Room at Perimeter Institute on 8 September 2026, from 15:30 to 17:00 Toronto time.

Large recurrent networks are important models in several fields, including neuroscience, machine learning, physics, and applied mathematics. Yet their dynamics are difficult to study directly, because high-dimensional nonlinear systems can exhibit rich behavior that is hard to summarize in terms of individual trajectories. In this talk, I will discuss an approach that seeks to understand such dynamics through the structure of the network’s equilibria. I will focus on a random balanced network of threshold-linear units that undergoes a transition from a single stable equilibrium to extensive chaos as the disorder strength crosses a critical value. Using a combination of Kac–Rice theory, replica calculations, numerical root-finding, and dynamical mean-field theory, we show that the chaotic regime contains an exponentially large number of equilibria. These equilibria are all saddles, but with only a fractionally small number of unstable directions. Surprisingly, despite the completely random connectivity, the equilibria are not scattered randomly through phase space. Instead, they are strongly correlated and confined to a comparatively small region. The chaotic attractor lies within this same region, suggesting a direct geometric link between the organization of unstable equilibria and the collective structure of the dynamics. This picture helps explain why networks with extensive chaos can nevertheless display dynamics dominated by a relatively small number of collective modes. More broadly, the results suggest that the geometry of equilibria provides a useful complementary perspective to dynamical mean-field theory for understanding high-dimensional neural dynamics. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2026-05-27. Recording duration: 00:46:40.

Traditional work in the study of human reward-based learning involves designing an experimental task---often inspired by Reinforcement Learning (RL) theory---and fits a small set of computational models---often inspired by RL algorithms---to that dataset. For example, researchers often model human behavior on bandit tasks using variants of Q-learning. While this approach has been highly productive, leading to landmark discoveries such as the dopamine reward prediction error hypothesis, it also has limitations. This talk focuses on the lack of generalizability of such models: Even if they closely fit behavior on the original task, models derived from the one-task-one-model paradigm usually predict behavior on other tasks quite poorly. I argue that this lack of generalizability is a fundamental problem for the cognitive sciences: we intuitively expect our models to be robust to superficial task differences, such as variations in the number of choice options, reward probabilities, or the exact kind of non-stationarity. I will propose potential solutions to this problem along two dimensions: the behavioral dataset and the computational model. Regarding computational models, I will introduce work in which we moved beyond the limitations of hand-crafted one-off models by employing flexible, data-driven methods. These methods allowed us to compare classes of models instead of individual model instances, allowing us to cover the space of possible models more exhaustively, and innovate cognitive mechanisms very efficiently. For the behavioral dataset, we move from using single learning tasks to a comprehensive task space that encompasses most existing paradigms in the literature, while closing the gaps between them in a near-continuous fashion. Our results suggest that more general models in conjunction with broader datasets can pave the road toward increasingly general models of human reward-based learning and decision making, and a persistent departure from many aspects of RL theory. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2026-05-20. Recording duration: 00:51:19.

Networks of interconnected neurons display diverse patterns of activity. Relating these patterns to the structure of the network is a central goal of theoretical neuroscience. Classic neural field and rate models have been powerful tools for this purpose due to their analytical tractability. Here, we show that the recently-developed combinatorial threshold-linear network (CTLN) model is a mean-field theory for excitatory-inhibitory Hawkes networks, with clustered connectivity, in an inhibition-stabilized regime. This mapping allows us to leverage powerful analytical results for CTLN networks to predict diverse macroscopic dynamics of clustered Hawkes networks, including metastability between various macroscopic fixed points, limit cycles, and chaotic attractors. We will then examine an extension of this approach to models with nonlinear dendritic dynamics, focusing on dendritic calcium spikes.We uncover a marked point process mean-field theory for these n etworks and use this to examine how somatic vs dendritic-targeting connectivity shapes the mean-field equilibrium phase diagram. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2026-05-13. Recording duration: 00:54:14.

Neural oscillations are often proposed to support brain computation by routing information, organizing cell assemblies, or shaping coding dynamics. Yet these ideas usually assume rhythms that are strong, sustained, and regular, whereas in vivo oscillations are often weak, transient, noisy, and variable in frequency and phase. In this talk, I will argue that such “no-metronome” oscillations are not just noisy fluctuations, but coordinated complex dynamics with functional consequences. Combining analyses of neural activity recordings during actual behavior (mice and non-human-primate LFPs and human EEG) with computational modelling, I will discuss evidence that transient oscillatory events can carry task-relevant information and support flexible communication through spatiotemporally structured relationships across populations, timescales, and frequencies. Together, these results suggest that oscillatory weakness and weirdness are not just imperfections, noise to average-out, but part of the functional repertoire of neural computation Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2026-05-06. Recording duration: 00:40:47.