Synaptic weight decay with selective consolidation enables fast learning without catastrophic forgetting

By Pascal Leimer, Michael H. Herzog, Walter Senn. 2019

12 citations recorded by the source corpus.

Learning can interfere with pre-existing memories that in classical neural networks may lead to catastrophic forgetting. Different from these networks, biological synapses show an early decay of long-term potentiation, combined with a tag & capture mechanism for selective consolidation. We present a 2-component synaptic plasticity model that, by an early decay and a tag & capture mechanism, enables context-dependent fast learning without catastrophic forgetting. For reinforcement learning in multiple contexts, the fast learning combined with an early weight decay maximizes the expected reward while minimizing interferences between subsequent contexts. Fast learning, enabled by a highly plastic weight component, improves performance for a given context. Between contexts this plastic component decays to prevent interference, but selective consolidation into a stable component protects old memories. As a downside of these mechanisms, learning is hampered when consolidation is triggered prematurely by interleaving easy and difficult tasks, consistent with human psychophysical experiments.

Subfields: Artificial Intelligence, Cognitive Neuroscience, Electrical and Electronic Engineering, Optimization & Training, Meta & Continual Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Cognitive Neuroscience

In Electrical and Electronic Engineering

In Optimization & Training

In Meta & Continual Learning

Citations in the corpus

Experience Replay for Continual Learning (2018, background): “This process complements local consolidation of past experience at the level of individual neurons and synapses, which is also believed to be present in biological neural networks [2, 14].”
Neural networks with late-phase weights (2020, background): “, 2017), and continual learning with biologically-realistic synapses (Kaplanis et al., 2018; Leimer et al., 2019), to name a few.”