Explainable AI Seminars @ Imperial

CLArg Group, Centre for eXplainable Artificial Intelligence, Imperial College London

XAI has witnessed unprecedented growth in both academia and industry in recent years (alongside AI itself), given its crucial role in supporting human-AI partnerships. XAI allows to deploy opaque data-driven AI methods safely in a variety of settings, such as finance, healthcare and law. XAI is positioned at the intersection of AI, human-computer interaction, the social sciences (in particular psychology) and applications.
Overall, XAI is increasingly part of all AI policies on ethics, trustworthiness and safety of AI. This seminar series focuses on all aspects of XAI, ranging from methods to applications.

Get invites in your inbox - Sign up here

Next Seminar

Noam Levi - Universality and Reductionism in Deep Learning

2^nd July, 15:00 BST, Imperial College London South Kensington Campus, Skempton 163 (directions) and on Microsoft Teams (join online).

Abstract

How is it that we can describe the complex universe using simple, fundamental rules? The equivalence between vastly different, complex physical systems, when observed from afar, enables us to make accurate predictions without analyzing their microscopic details. Conversely, by reducing such systems to their minimal constituents, we can uncover phenomena that would otherwise seem inscrutable. In this talk, I will explore how these concepts of universality and reductionism extend beyond the natural universe to the synthetic world of neural networks, focusing on a surprising phenomenon that can arise when training deep neural networks: Grokking, or delayed generalization. First, I will demonstrate that Grokking can occur even in seemingly trivial models without hidden layers or nonlinearities when they are near a “generalization threshold.” Next, I will show that these simple systems can be analyzed analytically, revealing that the Grokking phenomenon stems from the system approaching a critical point. Finally, I will discuss how Grokking might be understood as a manifestation of a broader characteristic of neural networks trained near criticality and outline prospects for future research.

Bio

Noam is a AI4Science Fellow at EPFL, affiliated with the Laboratory for Information and Inference Systems hosted by Prof. Volkan Cehver and the Laboratory of Astrophysics, hosted by Prof. Jean Paul Kneib. His research focuses on the intersection of theoretical physics and machine learning. Namely, applying tools from high energy and statistical physics to study machine learning algorithms as well as applying ML to advance the physical sciences. Noam received his double B.Sc. in Physics and Electrical Engineering in 2015, followed by his M.Sc. in 2017 and Ph.D. in Theoretical Particle Physics and Cosmology in 2024 from Tel Aviv University in Israel, under the supervision of Prof. Tomer Volansky.

Planned Seminars