# Synaptic facilitation and learning of multiplexed neural signals

## TL;DR

This paper introduces a new way for neural networks to encode information using the timing of spikes, allowing multiple signals to be processed efficiently in compact networks.

## Contribution

A novel synaptic plasticity mechanism enables learning of distinct temporal codings at the single-synapse level in spiking networks.

## Key findings

- Phase-coded spike trains can multiplex multiple information signals onto the same train, boosting information capacity.
- The learning rule allows synapses to adapt and specialize to different interspike intervals, enabling distinct temporal codings.
- The approach supports energy-efficient, compact networks for complex tasks using denser encoding.

## Abstract

In this work, we introduce a novel approach to one of the historically fundamental questions in neural networks: how to encode information? More particularly, we look at temporal coding in spiking networks, where the timing of a spike as opposed to the frequency, determines the information content. In contrast to previous temporal-coding schemes, which rely on the statistical properties of populations of neurons and connections, we employ a novel synaptic plasticity mechanism that allows the timing to be learnt at the single-synapse level.

Using a formal basis from information theory, we show how a phase-coded spike train (relative to some ‘reference’ phase) can, in fact, multiplex multiple different information signals onto the same spike train, significantly improving overall information capacity. We furthermore derive limits on the channel capacity in the phase-coded spiking case, and show that the learning rule also has a continuous derivative in the input-output relation, making it potentially amenable to classical learning rules from artificial neural networks such as backpropagation.

Using a simple demonstration network, we show the multiplexing of different signals onto the same connection, and demonstrate that different synapses indeed can adapt using this learning rule, to specialise to different interspike intervals (i.e., phase relationships). The overall approach allows for denser encoding, and thus energy efficiency, in neural networks for complex tasks, allowing smaller and more compact networks to achieve combinations of tasks which traditionally would have required high-dimensional embeddings.

Although carried out as a study in computational spiking neural networks, the results may have insights for functional neuroscience, and suggest links to mechanisms that have been shown from neuroscientific studies to support temporal coding. To the best of our knowledge, this is the first study to solve one of the outstanding problems in spiking neural networks: to demonstrate that distinct temporal codings can be distinguished through synaptic learning.