Seminars

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.

Robot missions require both collision-free trajectories and logical task ordering, such as collecting a resource before entering a restricted region. Augmented graphs of convex sets combine continuous trajectory optimization and discrete task sequencing in one problem. A layered graph constructed from an exact convex partition of free space selects an optimal task order and trajectory through a shortest-path calculation, exact up to finite Bézier parameterization. Its structure corresponds to Bellman–Held–Karp dynamic programming for the travelling-salesperson problem, retaining singly exponential worst-case complexity and improving exponentially on general temporal-logic tools. A library of specifications covers ordered collection, alternative keys, combined prerequisites, timing and conditional logic with correctness guarantees. Numerical benchmarks demonstrate exponential speedups and near-global optimality. Possible extensions include moving environments, safety-aware planning with conformal prediction sets and a large galactic-survey benchmark.

Florian Marquardt surveys approaches to reducing the energy and computational cost of neural networks by implementing their functions in physical devices. The central question is how experiments can efficiently determine updates to adjustable parameters, allowing the hardware itself to learn. Examples include learning through time reversal, obtaining nonlinear computation from linear optical scattering, and training analog quantum simulators as neuromorphic systems. The seminar is available in person and through the registration link on the organizer’s event page.

In this talk, I will explore how social affective biases arise even in the absence of motivational factors as an emergent outcome of the basic structure of social learning. In several studies, we found that initial negative interactions with some members of a group can cause subsequent avoidance of the entire group, and that this avoidance perpetuates stereotypes. Additional cognitive modeling discovered that approach and avoidance behavior based on biased beliefs not only influences the evaluative (positive or negative) impressions of group members, but also shapes the depth of the cognitive representations available to learn about individuals. In other words, people have richer cognitive representations of members of groups that are not avoided, akin to individualized vs group level categories. I will end presenting a series of multi-agent reinforcement learning simulations that demonstrate the emergence of these social-structural feedback loops in the development and maintenance of affective biases.

Nearest-neighbor indexes usually rely on a single distance function, but accurate comparisons can be expensive. This talk proposes a bi-metric framework: a cheap proxy metric builds the index, while the query procedure uses a limited number of evaluations of both the proxy and an expensive ground-truth metric. The theory applies to DiskANN and Cover Tree. When the proxy approximates the ground-truth metric within a bounded factor, the resulting structure can achieve arbitrarily good approximation guarantees under the accurate metric. Experiments on text retrieval using models with very different computational costs show improved accuracy-efficiency tradeoffs on almost all MTEB datasets compared with alternatives such as reranking. Joint work with Haike Xu and Sandeep Silwal.

Cognitive maps confer animals with flexible intelligence by representing spatial, temporal, and abstract relationships that can be used to shape thought, planning, and behavior. Cognitive maps have been observed in the hippocampus, but their algorithmic form and the processes by which they are learned remain obscure. Here, we employed large-scale, longitudinal two-photon calcium imaging to record activity from thousands of neurons in the CA1 region of the hippocampus while mice learned to efficiently collect rewards from two subtly different versions of linear tracks in virtual reality. The results provide a detailed view of the formation of a cognitive map in the hippocampus. Throughout learning, both the animal behavior and hippocampal neural activity progressed through multiple intermediate stages, gradually revealing improved task representation that mirrored improved behavioral efficiency. The learning process led to progressive decorrelations in initially similar hippocampal neural activity within and across tracks, ultimately resulting in orthogonalized representations resembling a state machine capturing the inherent struture of the task. We show that a Hidden Markov Model (HMM) and a biologically plausible recurrent neural network trained using Hebbian learning can both capture core aspects of the learning dynamics and the orthogonalized representational structure in neural activity. In contrast, we show that gradient-based learning of sequence models such as Long Short-Term Memory networks (LSTMs) and Transformers do not naturally produce such orthogonalized representations. We further demonstrate that mice exhibited adaptive behavior in novel task settings, with neural activity reflecting flexible deployment of the state machine. These findings shed light on the mathematical form of cognitive maps, the learning rules that sculpt them, and the algorithms that promote adaptive behavior in animals. The work thus charts a course toward a deeper understanding of biological intelligence and offers insights toward developing more robust learning algorithms in artificial intelligence.

Ophthalmology is ideally placed to benefit from recent advances in artificial intelligence. It is a highly image-based specialty and provides unique access to the microvascular circulation and the central nervous system. This talk will demonstrate diverse applications of machine learning and deep learning techniques in ophthalmology, including in age-related macular degeneration (AMD), the leading cause of blindness in industrialized countries, and cataract, the leading cause of blindness worldwide. This will include deep learning approaches to automated diagnosis, quantitative severity classification, and prognostic prediction of disease progression, both from images alone and accompanied by demographic and genetic information. The approaches discussed will include deep feature extraction, label transfer, and multi-modal, multi-task training. Cluster analysis, an unsupervised machine learning approach to data classification, will be demonstrated by its application to geographic atrophy in AMD, including exploration of genotype-phenotype relationships. Finally, mediation analysis will be discussed, with the aim of dissecting complex relationships between AMD disease features, genotype, and progression.

I will discuss work with Jack Lindsey modeling reinforcement learning for action selection in the basal ganglia. I will argue that the presence of multiple brain regions, in addition to the basal ganglia, that contribute to motor control motivates the need for an off-policy basal ganglia learning algorithm. I will then describe a biological implementation of such an algorithm that predicts tuning of dopamine neurons to a quantity we call "action surprise," in addition to reward prediction error. In the same model, an implementation of learning from a motor efference copy also predicts a novel solution to the problem of multiplexing feedforward and efference-related striatal activity. The solution exploits the difference between D1 and D2-expressing medium spiny neurons and leads to predictions about striatal dynamics.

Synaptic plasticity is known to be a key player in the brain’s life-long learning abilities. However, due to experimental limitations, the nature of the local changes at individual synapses and their link with emerging network-level computations remain unclear. I will present a numerical, meta-learning approach to deduce plasticity rules from either neuronal activity data and/or prior knowledge about the network's computation. I will first show how to recover known rules, given a human-designed loss function in rate networks, or directly from data, using an adversarial approach. Then I will present how to scale-up this approach to recurrent spiking networks using simulation-based inference.

Biologically plausible spiking neural networks (SNNs) are an emerging architecture for deep learning tasks due to their energy efficiency when implemented on neuromorphic hardware. However, many of the biological features are at best irrelevant and at worst counterproductive when evaluated in the context of task performance and suitability for neuromorphic hardware. In this talk, I will present an alternative paradigm to design deep learning architectures with good task performance in real-world benchmarks while maintaining all the advantages of SNNs. We do this by focusing on two main features – event-based computation and activity sparsity. Starting from the performant gated recurrent unit (GRU) deep learning architecture, we modify it to make it event-based and activity-sparse. The resulting event-based GRU (EGRU) is extremely efficient for both training and inference. At the same time, it achieves performance close to conventional deep learning architectures in challenging tasks such as language modelling, gesture recognition and sequential MNIST.

How could machines learn as efficiently as humans and animals? How could machines learn to reason and plan? How could machines learn representations of percepts and action plans at multiple levels of abstraction, enabling them to reason, predict, and plan at multiple time horizons? I will propose a possible path towards autonomous intelligent agents, based on a new modular cognitive architecture and a somewhat new self supervised training paradigm. The centerpiece of the proposed architecture is a configurable predictive world model that allows the agent to plan. Behavior and learning are driven by a set of differentiable intrinsic cost functions. The world model uses a new type of energy-based model architecture called H-JEPA (Hierarchical Joint Embedding Predictive Architecture). H-JEPA learns hierarchical abstract representations of the world that are simultaneously maximally informative and maximally predictable.

In the Learning Salon, we will discuss the similarities and differences between biological and machine learning, including individuals with diverse perspectives and backgrounds, so we can all learn from one another.

Over the course of a lifetime, we process a continual stream of information. Extracted from this stream, memories must be efficiently encoded and stored in an addressable manner for retrieval. To explore potential mechanisms, we consider a familiarity detection task where a subject reports whether an image has been previously encountered. We design a feedforward network endowed with synaptic plasticity and an addressing matrix, meta-learned to optimize familiarity detection over long intervals. We find that anti-Hebbian plasticity leads to better performance than Hebbian and replicates experimental results such as repetition suppression. A combinatorial addressing function emerges, selecting a unique neuron as an index into the synaptic memory matrix for storage or retrieval. Unlike previous models, this network operates continuously, and generalizes to intervals it has not been trained on. Our work suggests a biologically plausible mechanism for continual learning, and demonstrates an effective application of machine learning for neuroscience discovery.

In the Learning Salon, we will discuss the similarities and differences between biological and machine learning, including individuals with diverse perspectives and backgrounds, so we can all learn from one another.

In the Learning Salon, we will discuss the similarities and differences between biological and machine learning, including individuals with diverse perspectives and backgrounds, so we can all learn from one another.

A rich data representation that finds wide application in industry and research is the so-called knowledge graph - a graph-based structure where entities are depicted as nodes and relations between them as edges. Complex systems like molecules, social networks and industrial factory systems can be described using the common language of knowledge graphs, allowing the usage of graph embedding algorithms to make context-aware predictions in these information-packed environments.

Like last year this online workshop brings together researchers in the field to present their work and discuss ways of translating these findings into a better understanding of neural circuits. Topics include artificial and biologically plausible learning algorithms and the dissection of trained spiking circuits toward understanding neural processing. We have a manageable number of talks with ample time for discussions. This year’s executive committee comprises Chiara Bartolozzi, Sander Bohté, Dan Goodman, and Friedemann Zenke.

In this workshop you'll learn the fundamentals of PyTorch using an incremental, from-first-principles approach. We'll start with tensors, autograd, and the dynamic computation graph, and then move on to developing and training a simple model using PyTorch's model classes, datasets, data loaders, optimizers, and more. You should be comfortable using Python, Jupyter notebooks, Google Colab, Numpy and, preferably, object oriented programming.