An excitatory plasticity rule for the formation of neural clusters in populations of spiking neurons

Growing evidence suggests that metastable neural dynamics might be the substrate of important cortical computations. Metastable dynamics can be obtained in a spiking network model partitioned in clusters wherein each cluster of neurons has an average synaptic strength above a critical value. One crucial question is to understand how clusters, and hence metastable dynamics, can be formed by experience. Motivated by obtaining a minimal learning rule which relies only on local ingredients, we propose an excitatory plasticity rule combined with a homeostatic mechanism that keeps the neural activity close to a desired level. A synaptic weight change is triggered by the arrival of presynaptic spikes, while the polarity of change depends on the recent history of its synaptic inputs. This learning rule leads to stable and robust formation of clusters and metastability, which is then maintained in the face of endogenously generated ongoing activity. After training, the spiking network can be successfully retrained with a new set of stimuli. When neurons respond to multiple stimuli during training, the same learning rule leads to the formation of overlapping clusters, a necessary condition for storing an extensive number of memories. These results show how metastable neural dynamics can emerge from a biologically plausible plasticity rule without the need for non-local ingredients such as synaptic renormalization.