[PS-1.16] The computational mechanisms underlying learning during sleep

Schapiro, A. 1 , Herd, S. 2 , Trippe, A. 1 , O'Reilly, R. 2 , Rogers, T. 3 & Norman, K. 1

1 Department of Psychology and Princeton Neuroscience Institute, Princeton University
2 Department of Psychology and Neuroscience, University of Colorado-Boulder
3 Department of Psychology, University of Wisconsin-Madison

Sleep is thought to be crucial for the transfer of newly acquired information stored in the hippocampus to long term cortical memory systems. Very little is known, however, about how this new information is integrated into existing cortical knowledge structures. To accomplish this goal, existing cortical representations need to be adjusted to accommodate the new information without causing catastrophic loss of existing information. Furthermore, this delicate process needs to be accomplished while the brain is receiving minimal external inputs. Here, we describe a neural network model of how these goals could be accomplished during REM sleep. The model builds on prior work by Norman et al. (2005), showing how oscillating inhibition can be used to identify weak points in existing memories (so they can be strengthened) and to identify situations where stored memories are encroaching on one another (so they can be adaptively differentiated). During simulated REM sleep, the network visits attractors corresponding to stored memories, while synaptic depression causes the network to transition between attractors. The model learns by identifying stable activation states and treating them as "plus" (i.e., desirable) patterns; weights are updated by a Contrastive Hebbian Learning rule that compares neural coactivation at a synapse during the "plus" period to neural coactivation that is present while the pattern transitions out of this stable state (due to accumulating synaptic depression and inhibitory oscillations). We show that the simulated REM process significantly improves the model's ability to retain old knowledge while simultaneously incorporating newly-learned information. We also discuss extensions to the model (including mechanisms for "focusing" REM rehearsal on regions of memory space that were most impacted by new learning) and new applications of the model to experiments that explore the effects of sleep on category learning.