Research Fellow Candidate Presentations

Please hold this time for presentations from our 2025-2026 Research Fellow candidates.

9:00-9:45am – L’emir Omar Chehab, ENSAE Paris

Understanding the success of modern energy-based modelling

• Energy-based probabilistic models are enjoying renewed popularity, given their success in applications to language and image processing. Modern methods for estimating and sampling these models overwhelmingly rely on annealing, which employs a sequence of intermediate distributions between the Gaussian and the data distributions. My research aims to identify the limitations of older methods, quantify how annealing improves their efficiency, and design principled strategies for selecting the intermediate distributions.

9:45-10:30am – David Clark, Columbia University

Theories of Dimensionality, Dynamics, and Computation in Large Recurrent Neural Networks

• Neural circuits are characterized by their large numbers of neurons, nonlinear dynamics, recurrent connectivity, and synaptic connections that change across multiple timescales. These features, while fundamental to neural computations underlying sensation, cognition, and behavior, pose significant challenges for theorists. My research addresses these challenges using tools from physics and machine learning. In this talk, I will present theories of neural-network activity, ranging from high-dimensional chaos to low-dimensional continuous attractors, elucidating principles of collective dynamics and information processing relevant to both biological and artificial neural systems.

10:30-11:15am – Alexandru Damian, Princeton University

Foundations of Deep Learning Optimization

• Classical optimization theory fails to capture many deep learning phenomena, leading to an over-reliance on heuristics and trial-and-error to make critical training decisions. A key challenge is that deep learning optimizers exhibit complex dynamics, including oscillations, feedback loops, and chaos, which make their behavior difficult to predict or analyze. My research in optimization theory characterizes the effects of these dynamics in order to make an optimizer’s behavior explicit and interpretable. This work has revealed that many optimization behaviors crucial to deep learning’s success emerge implicitly during training, rather than being explicitly programmed. These emergent behaviors enable optimizers to find solutions that generalize well, remain stable in challenging loss landscapes, and efficiently adapt to local curvature. I will highlight a few of these results and discuss some future directions which aim to leverage these insights for efficient hyperparameter tuning, optimizer design, and large scale training.

11:15am-12:00pm – Clementine Domine, University College London (UCL)

Exact learning dynamics of the rich and lazy learning regimes

• Previous research has demonstrated that networks can learn through distinct learning regimes: a “lazy” regime, where internal representations remain largely unchanged during training, and a “rich” or “feature-learning” regime, where representations dynamically adapt to the task. This work investigates the key initialization factors —specifically the relative scale of weights across layers – that govern the learning regime in neural networks. We propose a model that characterizes the continuum between lazy and rich learning, providing exact analytical solutions to the learning dynamics with practical implications for applications in both neuroscience and machine learning.