Can sleep protect memories from catastrophic forgetting?

By Oscar C. González, Yury Sokolov, Giri Panamoottil Krishnan, Maxim Bazhenov. 2019

64 citations recorded by the source corpus.

Continual learning remains to be an unsolved problem in artificial neural networks. Biological systems have evolved mechanisms by which they can prevent catastrophic forgetting of old knowledge during new training and allow lifelong learning. Building upon data suggesting the importance of sleep in learning and memory, here we test a hypothesis that sleep protects memories from catastrophic forgetting. We found that training in a thalamocortical network model of a “new” memory that interferes with previously stored “old” memory may result in degradation and forgetting of the old memory trace. Simulating NREM sleep immediately after new learning leads to replay, which reverses the damage and ultimately enhances both old and new memory traces. Surprisingly, we found that sleep replay goes beyond recovering old memory traces that were damaged by new learning. When a new memory competes for the neuronal/synaptic resources previously allocated to the old memory, sleep replay changes the synaptic footprint of the old memory trace to allow for the overlapping populations of neurons to store multiple memories. Different neurons become preferentially supporting different memory traces to allow successful recall. We compared synaptic weight dynamics during sleep replay with that during interleaved training – a common approach to overcome catastrophic forgetting in artificial networks – and found that interleaved training promotes synaptic competition and weakening of reciprocal synapses, effectively reducing an ensemble of neurons contributing to memory recall. This leads to suboptimal recall performance compared to that after sleep. Together, our results suggest that sleep provides a powerful mechanism to achieve continual learning by combining consolidation of new memory traces with reconsolidation of old memory traces to minimize memory interference.

Subfields: Cognitive Neuroscience, Electrical and Electronic Engineering

How researchers see this paper

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

In Cognitive Neuroscience

In Electrical and Electronic Engineering

Citations in the corpus

Replay in Deep Learning: Current Approaches and Missing Biological Elements (2021, neutral): “Instead, sleep may change representations of the old memories by reassigning different subsets of neurons and synapses to effectively orthogonalize memory representations and allow for overlapping populations of neurons to store multiple competing memories (González et al., 2020).”
Biologically inspired sleep algorithm for artificial neural networks (2019, background): “Effect of sleep: Sleep can increase difference among weights (which are different enough to begin with) as was shown in [30, 31].”
Biologically Inspired Sleep Algorithm for Reducing Catastrophic Forgetting in Neural Networks (2020, background): “Biophysical modelling of sleep has shown that sleep can recover memories lost due to catastrophic forgetting by replaying old memories and decreasing representational overlap among interfering memories (Gonzalez et al. 2019).”
Sleep prevents catastrophic forgetting in spiking neural networks by forming a joint synaptic weight representation (2022, background): “We previously showed that sleep replay improves memory in a thalamocortical network [38–40] and when a network was trained to learn interfering tasks sequentially, sleep prevented the old task memory from catastrophic forgetting [41].”
Biologically inspired sleep algorithm for increased generalization and adversarial robustness in deep neural networks (2020, methodology): “These changes entail both an increase in synaptic weights associated with a specific task and pruning of synapses involved in other tasks (Gonzalez et al., 2019; Tononi & Cirelli, 2006).”