Séminaire du CIMMUL – Guillaume Lajoie

Rich and Lazy neurons: network connectivity structure and the double implications of feature learning for generalization

Guillaume Lajoie

Canada Research Chair in Neural Computation and Interfacing
Canada CIFAR AI chair
Maths & Stats Dept — Univ de Montréal — Mila, Quebec AI Institute

Résumé

In this talk, I will present an overview of recent advancements in deep learning theory that characterize learning dynamics in large network parameter spaces. Specifically, I will explore the question of learning regimes characterized by the movement of the Neural Tangent Kernel during optimization. First, I will discuss how network connectivity can influence learning regimes, and how task requirements together with synaptic weight distributions influence the amount of changes one can expect in synaptic weights during a task acquisition. Second, I will explore the implication of distinct learning regimes on generalization performance of neural networks.

Bio: Guillaume Lajoie is an Associate Professor at the Dept. of Mathematics and Statistics at the Université de Montréal. He holds a Canada CIFAR AI Research Chair and a Canada Research Chair in Neural Computation and Interfacing. His research group works at the intersection of AI and Neuroscience, developing mathematical tools to better understand neural networks, biological or artificial, as well as algorithms for brain-machine interfaces for scientific and clinical use

Le séminaire aura lieu au local 3820 du pavillon Alexandre-Vachon et en ligne.

Pour rejoindre la réunion Zoom :
https://ulaval.zoom.us/j/62680136430?pwd=eldBYjdNTG5QR2VxTTFqbVM4UGVRZz09

Meeting ID: 626 8013 6430
Passcode: 693150