35 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Towards Biologically Plausible Deep Learning: Early Inference in Energy-Based Models Approximates Back-Propagation
University of Bonn
Recording
Abstract
Asja Fischer describes how early Langevin inference steps in an energy-based latent-variable model approximate the propagation of error signals through hidden layers. In this account, changes in neural activity over time supply a learning signal for a local weight update. Theory and simulations connect the update to spike-timing-dependent plasticity and suggest a way for inference and credit assignment to share neural dynamics.
Topics
- credit assignment
- energy-based models
- spike-timing-dependent plasticity
- Biologically Plausible Deep Learning, broader topic: deep learning
- early Langevin inference, broader topic: inference
- latent variable model
- error signals, search World Wide
- hidden layers, search World Wide
- neural activity
- local weight update, search World Wide