Bayesian Metaplasticity from Synaptic Uncertainty

Djohan Bonnet, Tifenn HIRTZLIN, Tarcisius Januel, Thomas Dalgaty, Damien Querlioz, Elisa Vianello

Keywords: Bayesian neural network, biologically-plausible machine-learning, metaplasticity

TL;DR: In this study, we introduce MEtaplasticity from Synaptic Uncertainty (MESU),we demonstrate MESU's remarkable capability to maintain learning performance across 100 tasks without the need of explicit task boundaries.

Abstract: Catastrophic forgetting remains a challenge for neural networks, especially in lifelong learning scenarios. In this study, we introduce MEtaplasticity from Synaptic Uncertainty (MESU), inspired by metaplasticity and Bayesian inference principles. MESU harnesses synaptic uncertainty to retain information over time, with its update rule closely approximating the diagonal Newton's method for synaptic updates. Through continual learning experiments on permuted MNIST tasks, we demonstrate MESU's remarkable capability to maintain learning performance across 100 tasks without the need of explicit task boundaries.

Submission Number: 8