Learning and Variability in Spiking Neural Networks

Neural networks exhibit ongoing, spatiotemporal patterns of spiking activity. Evidence shows that these patterns are metastable, i.e. temporary, transient, and non-stationary. Metastability is theorized to be adaptive for neural and cognitive function, but learning must somehow remain stable in the context of highly variable spike dynamics. In the present study, a neural network learning algorithm is developed to co-exist with intrinsic variability that arises from regulating spike propagation to stay near its critical branching point. The learning algorithm is based on reinforcement traces stored at synapses that change much more slowly than synaptic switches triggered to maintain critical branching. As a result, learning establishes a stable synaptic space within which variability and metastability can arise from critical branching. Model efficacy is demonstrated using timedelayed XOR learning, and spike dynamics are compared with evidence of metastability in hippocampal recordings.