IBM Israel Research Seminars

This work is aimed at deriving Reinforcement Learning algorithms for biologically based networks of neurons. We focus on applications of gradient based policy learning to networks of spiking neurons, with spike time dependent plasticity taking place at the synaptic connections between neurons. We present a basic learning algorithm for stochastic Integrate and Fire neurons, and extend it to networks of generalized neural elements. Applications of the algorithm to several learning tasks are presented. Finally, we present an analysis relating the derived rules to well known biological plasticity rules. The general approach allows us to interpret physiologically motivated local synaptic update rules through global optimality principles.