Abstract

The brain’s ability to retain memories over a lifespan is thought to rely on the reactivation of memory-representing cell assemblies. To experimentally test this, for my PhD I performed multi-unit recordings in the hippocampus of mice exploring novel environments. We found that selectively disrupting reactivation by closed-loop optogenetic silencing of sharp-wave/ripples impaired the later reinstatement of recently-formed place cell assemblies, thereby providing the first direct evidence that reactivation stabilizes new cell assemblies. To gain deeper insight into the computational role of reactivation, for my postdoc I turned to machine learning. Current state-of-the-art artificial neural networks are my ideal “model organism” as they can perform extremely well on a wide variety of individual tasks, but they struggle to retain old information when trained on new tasks. Could reactivation improve memory consolidation in artificial neural networks? To test this, we equipped deep feed-forward networks for classification with feedback connections trained to have generative capability. We found that interleaving “reactivation” generated by these feedback connections with new task data substantially reduced the catastrophic forgetting of old tasks. Notably, for classification of MNIST -digits, this approach outperforms and is more widely applicable then current deep learning strategies for alleviating catastrophic forgetting. To finish, I will discuss (1) why I believe this approach could scale to more complicated inputs and (2) how I plan to use further insights from the brain to achieve this.