35 min
Seminar, Hosted by Montreal AI and Neuroscience (MAIN) · UNIQUE Centre
Credit assignment in neural networks without plasticity
Université de Montréal & Mila
Recording
Abstract
Guillaume Lajoie examines how neural networks can assign credit and adapt through input-driven changes in activity while their parameters stay fixed. He considers how predictive training sets up this capacity, how expressive such non-parametric learning can be and how different meta-training objectives shape it. The approach offers a common way to study fast adaptation in biological networks and in-context learning in pretrained AI models.
Topics
- credit assignment
- in-context learning
- neural networks
- input-driven changes in activity, search World Wide
- predictive training, broader topic: training
- non-parametric learning, broader topic: learning
- meta-training objectives, search World Wide
- fast adaptation, broader topic: adaptation
- pretrained AI models, broader topic: AI models