Time Divergence-Convergence Learning Scheme in Multi-Layer Dynamic Synapse Neural Networks

TL;DR

This paper introduces a novel time divergence-convergence (TDC) learning scheme for multi-layer dynamic synapse neural networks, enabling effective spike train classification and sub-second temporal processing.

Contribution

The paper proposes the TDC learning scheme for DSNNs, improving training scalability and performance in spike train classification tasks.

Findings

Achieved over 92% accuracy in spike train classification.

Outperformed single-layer DSNN by 22%.

Demonstrated scalability for multi-layer DSNN training.

Abstract

A new learning scheme called time divergence-convergence (TDC) is proposed for two-layer dynamic synapse neural networks (DSNN). DSNN is an artificial neural network model, in which the synaptic transmission is modeled by a dynamic process and the information between neurons are transmitted through spike timing. In TDC, the intra-layer neurons of a DSNN are trained to map input spike trains to a higher dimension of spike trains called a feature-domain, and the output neurons are trained to build the desired spike trains by processing the spike timing of intralayer neurons. The DSNN performance was examined in a jittered spike train classification task which shows more than 92\% accuracy in classifying different spike trains. The DSNN performance is comparable with the recurrent multi-layer neural networks and surpasses a single-layer DSNN with a 22\% margin. Synaptic dynamics have been proposed as the neural substrate for sub-second temporal processing; we can utilize TDC to train a DSNN to perform diverse forms of sub-second temporal processing. The TDC learning proposed here is scalable in terms of the synaptic adaptation of deeper layers of multi-layer DSNNs. The DSNN along with TDC learning proposed here can be used in to replicate the processing observed in neural circuitry and in pattern recognition tasks.