The combination of Hebbian and predictive plasticity learns invariant object representations in deep sensory networks

Research salon

Manu Srinath Halvagal, Friedemann Zenke

https://www.biorxiv.org/content/10.1101/2022.03.17.484712v1

This paper is about

a self-supervised learning rule that

  • has a predictive component for temporally contiguous features
  • has a Hebbian component that prevents activity collapse

Ideas based on...

VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning

https://arxiv.org/abs/2105.04906

Adrien Bardes, Jean Ponce and Yann LeCun

'cake analogy' by LeCun, NIPS 2016 / ISSCC 2019

self-supervised learning by LeCun at ISSCC 2019

Bienenstock Cooper Munro (BCM) theory

Latent Predictive Learning (LPL)

learning rule

Self supervised learning in a DNN

Local layer learning vs. end-to-end

Summary

  • + self-supervised learning using an apparently biologically plausible rule
  • + does not need contrastive examples
  • + possibility to link this work to spike timing
  • - limited to toy problems
  • - no spiking in this work

LPL research salon

By Gregor Lenz

LPL research salon

  • 190