Two dimensional model of learning in spiking neural networks with homeostasis and reward

Abstract

The huge complexity of molecular mechanisms that support memory formation makes it difficult to build simple, but precise and sufficient models for an efficient simulation of large neural networks. In this paper, we propose the phenomenological model of a learning rule that describes the synaptic strength via slow and fast variables. Two variables interact with each other in a bidirectional manner that allows to combine the reward and unsupervised learning. Results show the stability of synaptic strength due to coupling of two variables and fast homeostatic plasticity. The multiplicative approach of synaptic scaling preserves memory patterns of statistically more frequent input signals. Similar to the eligibility traces approach, the model tracks recent synaptic changes and allows to reinforce these changes. Also, we speculate on a possible biophysical interpretation of such a model that includes the fast movement of receptors to the membrane and their stabilization into clusters.