Abstract
A central question in neuroscience is how to develop realistic models thatpredict output firing behavior based on provided external stimulus. Given a setof external inputs and a set of output spike trains, the objective is todiscover a network structure which can accomplish the transformation asaccurately as possible. Due to the difficulty of this problem in its mostgeneral form, approximations have been made in previous work. Pastapproximations have sacrificed network size, recurrence, allowed spiked count,or have imposed layered network structure. Here we present a learning rulewithout these sacrifices, which produces a weight matrix of a leakyintegrate-and-fire (LIF) network to match the output activity of bothdeterministic LIF networks as well as probabilistic integrate-and-fire (PIF)networks. Inspired by synaptic scaling, our pre-synaptic pool modification(PSPM) algorithm outputs deterministic, fully recurrent spiking neural networksthat can provide a novel generative model for given spike trains. Similarity inoutput spike trains is evaluated with a variety of metrics including avan-Rossum like measure and a numerical comparison of inter-spike intervaldistributions. Application of our algorithm to randomly generated networksimproves similarity to the reference spike trains on both of these statedmeasures. In addition, we generated LIF networks that operate near criticalitywhen trained on critical PIF outputs. Our results establish that learning rulesbased on synaptic homeostasis can be used to represent input-outputrelationships in fully recurrent spiking neural networks.