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.

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.

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 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.

Behavioural measurements are often overlooked by computational neuroscientists, who prefer to focus on electrophysiological recordings or neuroimaging data. This attitude is largely due to perceived lack of depth/richness in relation to behavioural datasets. I will show how contemporary psychophysics can deliver extremely rich and highly constraining datasets that naturally interface with computational modelling. More specifically, I will demonstrate how psychophysics can be used to guide/constrain/refine computational models, and how models can be exploited to design/motivate/interpret psychophysical experiments. Examples will span a wide range of topics (from feature detection to natural scene understanding) and methodologies (from cascade models to deep learning architectures).

Rajanlab designs neural network models constrained by experimental data, and reverse engineers them to figure out how brain circuits function in health and disease. Recently, we have been developing a powerful new theory-based framework for “in-vivo tract tracing” from multi-regional neural activity collected experimentally. We call this framework CURrent-Based Decomposition (CURBD). CURBD employs recurrent neural networks (RNNs) directly constrained, from the outset, by time series measurements acquired experimentally, such as Ca2+ imaging or electrophysiological data. Once trained, these data-constrained RNNs let us infer matrices quantifying the interactions between all pairs of modeled units. Such model-derived “directed interaction matrices” can then be used to separately compute excitatory and inhibitory input currents that drive a given neuron from all other neurons. Therefore different current sources can be de-mixed – either within the same region or from other regions, potentially brain-wide – which collectively give rise to the population dynamics observed experimentally. Source de-mixed currents obtained through CURBD allow an unprecedented view into multi-region mechanisms inaccessible from measurements alone. We have applied this method successfully to several types of neural data from our experimental collaborators, e.g., zebrafish (Deisseroth lab, Stanford), mice (Harvey lab, Harvard), monkeys (Rudebeck lab, Sinai), and humans (Rutishauser lab, Cedars Sinai), where we have discovered both directed interactions brain wide and inter-area currents during different types of behaviors. With this powerful framework based on data-constrained multi-region RNNs and CURrent Based Decomposition (CURBD), we ask if there are conserved multi-region mechanisms across different species, as well as identify key divergences.

The potential of machine learning and deep learning to advance artificial intelligence is driving a quest to build dedicated computers, such as neuromorphic hardware that emulate the biological processes of the brain. While the hardware technologies already exist, their application to real-world tasks is hindered by the lack of suitable programming methods. Advances at the interface of neural computation and machine learning showed that key aspects of deep learning models and tools can be transferred to biologically plausible neural circuits. Building on these advances, I will show that differentiable programming can address many challenges of programming spiking neural networks for solving real-world tasks, and help devise novel continual and local learning algorithms. In turn, these new algorithms pave the road towards systematically synthesizing machine intelligence in neuromorphic hardware without detailed knowledge of the hardware circuits.

At extreme scale, even conventional sampling and projection methods can become too expensive. This talk treats locality-sensitive hashing as an amortized constant-time adaptive sampler, connecting probabilistic hash tables with efficient unbiased statistical estimators. A few hash lookups can support adaptive estimation at a cost close to uniform sampling. Applications include partition-function estimation for large natural-language models such as word2vec, adaptive gradient estimation for stochastic gradient descent, and sublinear deep learning with very large parameter spaces. The talk also considers memory reduction through a hashing method related to count-min sketches. An example trains a classifier with 100,000 classes and 400,000 features on a single Titan X while using at most five percent of the memory needed to store all weights; a conventional logistic-regression model for that dataset requires 320 GB.

The success of deep neural networks raises questions about the scalability of kernel methods and the roles of large datasets, depth and training algorithms. This talk examines techniques that make kernel learning practical for large datasets in both scalar and multivariate prediction. The methods combine randomized data embeddings, Quasi-Monte Carlo acceleration, distributed convex optimization and input-output kernel learning. Experiments on speech-recognition and computer-vision datasets compare randomized kernel methods with deep neural networks and report essentially matching performance. The talk explores how randomized kernel constructions can resemble neural-network architectures, and how invariant kernel learning and matrix-valued kernels might support deeper architectures. It discusses research results and connections between these approaches.