Firing Without Wiring: How REM-like Sleep Reactivation Rebalances Excitation and Inhibition to Orthogonalize Memories

REM-like sleep plays a crucial role in memory consolidation, though its exact mechanisms have remained unclear. Various theories suggest REM sleep helps separate memories, creating distinct representations of similar ones; however, concrete evidence for this remains limited. To explore this, researchers employed recurrent neural networks trained with Equilibrium Propagation (EP), a biologically plausible learning algorithm, to study the impact of REM-like sleep on memory consolidation.

The findings revealed that REM-like sleep decreases memory interference by decoupling task-specific neuronal populations, thereby isolating individual memory traces. During sleep, reactivation of memory traces was predominantly driven by the most recently encoded memory. Crucially, net excitatory input onto this recently activated population decreased, with a significant reduction in input from older memory populations.

This selective reduction of excitatory input allowed the older memories to regain their own neuronal activity. The researchers found that distributing the same total synaptic change evenly across all memory populations would not produce the same recovery effect. They termed this phenomenon "reactivation-targeted pruning," where plasticity is restricted to a single population, causing neurons to lose inputs that are inactive when those neurons fire.

Even without inhibitory plasticity, as long as excitatory synapses were depressed, this mechanism held true. These results not only shed light on the potential mechanisms behind REM sleep's role in memory consolidation but also open the door to incorporating human-like learning behaviors into artificial neural networks.