Unraveling the role of collective bursting neurons, quiet waking, and structural plasticity in memory consolidation using a computational approach

When memorizing new information, it is commonly accepted that breaks associated with brain rest can improve performance. We investigate this hypothesis using a computational approach. Our neural network, composed of conductance-based model neurons, simulates brain states, transitioning from active learning to quiet waking. It corresponds to a neuronal switch from tonic firing to collective bursting orchestrated by neuromodulators. Simultaneously, the network modifies synaptic weights through plasticity to encode new memories.

Recent findings reveal a homeostatic reset induced by collective bursting across various traditional synaptic plasticity rules (pair-based, triplet, or calcium-based rule). Unintuitively, strong weights depress, and weak weights potentiate during bursting until a set point is reached, causing forgetting but also restoring synaptic weights and facilitating new memory formation. We propose a structural plasticity rule that complements traditional synaptic plasticity rules governing early-stage Long-Term Potentiation (E-LTP) and provides insights into late-stage Long-Term Potentiation (L-LTP).

In our study, we demonstrate the efficacy of this novel mechanism across diverse memory tasks. Initially, we observe that quiet waking underlying collective bursting enhances the Signal-to-Noise Ratio in a pairing memory task. Moving on to a pattern recognition task, the network adeptly learns to identify small patterns, whether overlapping or not. We thoroughly analyze the evolution of receptive fields, represented by pattern-associated weight matrices, during switches from active learning to quiet waking states. Remarkably, during quiet waking periods, memory consolidation occurs without any pattern recall. Extending this approach to the MNIST recognition task leads to notable improvements in performance. In all tasks, blocking quiet waking states decreases the ability to consolidate memory.

In conclusion, combining quiet waking with bursting neurons and structural plasticity improves learning and memory consolidation. This research aims to inspire investigations into the biophysical mechanisms of quiet waking in memory and the potential integration of resting states in machine learning algorithms for artificial intelligence.