Learning First-to-Spike Policies for Neuromorphic Control Using Policy Gradients

Artificial Neural Networks (ANNs) are currently being used as function approximators in many state-of-the-art Reinforcement Learning (RL) algorithms. Spiking Neural Networks (SNNs) have been shown to drastically reduce the energy consumption of ANNs by encoding information in sparse temporal binary spike streams, hence emulating the communication mechanism of biological neurons. Due to their low energy consumption, SNNs are considered to be important candidates as co-processors to be implemented in mobile devices. In this work, the use of SNNs as stochastic policies is explored under an energy-efficient first-to-spike action rule, whereby the action taken by the RL agent is determined by the occurrence of the first spike among the output neurons. A policy gradient-based algorithm is derived considering a Generalized Linear Model (GLM) for spiking neurons. Experimental results demonstrate the capability of online trained SNNs as stochastic policies to gracefully trade energy consumption, as measured by the number of spikes, and control performance. Significant gains are shown as compared to the standard approach of converting an offline trained ANN into an SNN.

I. Introduction

Artificial neural networks (ANNs) are used as parameterized non-linear models that serve as inductive bias for a large number of machine learning tasks, including notable applications of Reinforcement Learning (RL) to control problems [1]. While ANNs rely on clocked floating- or fixed-point operations on real numbers, Spiking Neural Networks (SNNs) operate in an event-driven fashion on spiking synaptic signals (see Fig. 1). Due to their lower energy consumption when implemented on specialized hardware, SNNs are emerging as an important alternative to ANNs that is backed by major technology companies, including IBM and Intel [2], [3]. Specifically, SNNs are considered to be important candidates as co-processors to be implemented in battery-limited mobile devices (see, e.g., [4]). Applications of SNNs, and of associated neuromorphic hardware, to supervised, unsupervised, and RL problems have been reported in a number of works, first in the computational neuroscience literature and more recently in the context of machine learning [5]–[7]. Fig. 1.

(a) SNN first-to-spike policy with action selected (r in the illustration) among up, down, left, and right marked with a bold line and decision time marked with a dashed vertical line; (b) an example of a realization of an action sequence in a windy grid-world problem.