Homeostatic Structural Plasticity and the Formation of Memory Engrams
Neurons in the brain use plastic synaptic connections to encode memories. This plasticity can manifest itself as a change of synaptic strength, but also by creating new and pruning existing connections. We derive a mean-field theory for dynamical networks of integrate-and-fire neurons which exhibit structural plasticity based on firing rate homeostasis. We analyze the process of memory formation, using this new theory in conjunction with simulations of recurrent networks. The networks are able to store repeatedly presented patterns, and to successfully recall patterns based on incomplete cues. Our analysis shows that the temporal dynamics of memorizing and forgetting can be decomposed into a fast homeostatic drift and a slow diffusion process. The presence of a stimulus effectively changes the nature of the homeostatic rule into a Hebbian rule that governs the formation of memory engrams. In fact, homeostatic plasticity induces a special type of "silent memories" and sheds light on the relation between homeostatic and Hebbian plasticity.