Digital Multiplier-less Event-Driven Spiking Neural Network Architecture for Learning a Context-Dependent Task

TL;DR

This paper presents a novel digital, multiplier-less spiking neural network architecture capable of learning context-dependent tasks via reinforcement learning, demonstrated through hardware implementation and robotic experiments.

Contribution

It introduces a new hardware-efficient SNN design that enables reinforcement learning for context-dependent tasks in a neuromorphic system.

Findings

Hardware implementation successfully learned stimulus-response associations

The system operates efficiently with low power and space requirements

Learning was demonstrated in a robot within a closed sensorimotor loop

Abstract

Neuromorphic engineers aim to develop event-based spiking neural networks (SNNs) in hardware. These SNNs closer resemble dynamics of biological neurons than todays' artificial neural networks and achieve higher efficiency thanks to the event-based, asynchronous nature of processing. Learning in SNNs is more challenging, however. Since conventional supervised learning methods cannot be ported on SNNs due to the non-differentiable event-based nature of their activation, learning in SNNs is currently an active research topic. Reinforcement learning (RL) is particularly promising method for neuromorphic implementation, especially in the field of autonomous agents' control, and is in focus of this work. In particular, in this paper we propose a new digital multiplier-less hardware implementation of an SNN. We show how this network can learn stimulus-response associations in a context-dependent task through a RL mechanism. The task is inspired by biological experiments used to study RL in animals. The architecture is described using the standard digital design flow and uses power- and space-efficient cores. We implement the behavioral experiments using a robot, to show that learning in hardware also works in a closed sensorimotor loop.