Stochastic variational learning in recurrent spiking networks.

Danilo Jimenez Rezende,Wulfram Gerstner

doi:10.3389/fncom.2014.00038

Abstract

The ability to learn and perform statistical inference with biologically plausible recurrent networks of spiking neurons is an important step toward understanding perception and reasoning. Here we derive and investigate a new learning rule for recurrent spiking networks with hidden neurons, combining principles from variational learning and reinforcement learning. Our network defines a generative model over spike train histories and the derived learning rule has the form of a local Spike Timing Dependent Plasticity rule modulated by global factors (neuromodulators) conveying information about “novelty” on a statistically rigorous ground. Simulations show that our model is able to learn both stationary and non-stationary patterns of spike trains. We also propose one experiment that could potentially be performed with animals in order to test the dynamics of the predicted novelty signal.

Highlights

Humans and animals are able to learn complex behavioral tasks and memorize events or temporally structured episodes
Plasticity (STDP) has intrigued theoreticians, because it provides a local Hebbian learning rule for spiking neurons; local, here, means that the dynamics of the synapses is of the form d dt wij h(posti, prej), where prej is the set of pre-synaptic variables of neuron j and posti is the set of post-synaptic variables of neuron i and h is an arbitrary functional
We have proposed an alternative to the learning algorithms previously proposed in Brea et al (2011) and Jimenez Rezende et al (2011) for learning a generative model of spike trains defined by recurrent spiking networks

Summary

Introduction

Humans and animals are able to learn complex behavioral tasks and memorize events or temporally structured episodes. With relation to neuroscience, unsupervised learning is most commonly related to developmental plasticity (Miller et al., 1989), formation of receptive fields (Song and Abbott, 2001) or cortical rewiring (Young et al, 2007). Most early applications of unsupervised STDP concern the learning of feedforward connections and the formation of receptive fields (Gerstner et al, 1996; Kempter et al, 1999; Song et al, 2000; Song and Abbott, 2001). Unsupervised STDP will tune to the earliest spikes (Song and Abbott, 2001; Gerstner and Kistler, 2002; Guyonneau et al, 2005) and can perform Independent Component Analysis (Clopath et al, 2010; Savin et al, 2010)

Methods

Results

Conclusion