TopSpark: A Timestep Optimization Methodology for Energy-Efficient Spiking Neural Networks on Autonomous Mobile Agents

TL;DR

TopSpark addresses the energy efficiency challenge of Spiking Neural Networks (SNNs) for autonomous mobile agents by introducing adaptive timestep optimization that spans both training and inference. It analyzes accuracy across timesteps, identifies influential neuron parameters, and applies lightweight parameter-enhancement policies along with a multi-objective trade-off to balance accuracy, latency, and energy. The approach achieves substantial latency reductions (~3.9x) and energy savings (~3.3x–3.5x) with accuracy losses typically below 2% across diverse networks and workloads, enabling practical online learning and runtime adaptation. This work advances the deployment of energy-aware neuromorphic systems on battery-powered robots and provides a framework for adaptive accuracy-energy-latency management in SNNs.

Abstract

Autonomous mobile agents require low-power/energy-efficient machine learning (ML) algorithms to complete their ML-based tasks while adapting to diverse environments, as mobile agents are usually powered by batteries. These requirements can be fulfilled by Spiking Neural Networks (SNNs) as they offer low power/energy processing due to their sparse computations and efficient online learning with bio-inspired learning mechanisms for adapting to different environments. Recent works studied that the energy consumption of SNNs can be optimized by reducing the computation time of each neuron for processing a sequence of spikes (timestep). However, state-of-the-art techniques rely on intensive design searches to determine fixed timestep settings for only inference, thereby hindering the SNNs from achieving further energy efficiency gains in both training and inference. These techniques also restrict the SNNs from performing efficient online learning at run time. Toward this, we propose TopSpark, a novel methodology that leverages adaptive timestep reduction to enable energy-efficient SNN processing in both training and inference, while keeping its accuracy close to the accuracy of SNNs without timestep reduction. The ideas of TopSpark include: analyzing the impact of different timesteps on the accuracy; identifying neuron parameters that have a significant impact on accuracy in different timesteps; employing parameter enhancements that make SNNs effectively perform learning and inference using less spiking activity; and developing a strategy to trade-off accuracy, latency, and energy to meet the design requirements. The results show that, TopSpark saves the SNN latency by 3.9x as well as energy consumption by 3.5x (training) and 3.3x (inference) on average, across different network sizes, learning rules, and workloads, while maintaining the accuracy within 2% of SNNs without timestep reduction.

Figures (11)

• Figure 1: (a) Autonomous mobile agents typically require low-power processing and capabilities for adapting to different operational environments to complete ML-based tasks. (b) SNN architecture that supports STDP-based unsupervised learning, i.e., fully-connected network, which is suitable for enabling ultra-low power/energy processing and efficient online learning for autonomous mobile agents. A larger SNN model has a higher number of excitatory neurons to recognize more features than a smaller one.
• Figure 2: Experimental results considering an SNN with 400 excitatory neurons with a fully-connected architecture in Fig. \ref{['Fig_SNNcaseNarch']}(b), rate coding, and pair-based STDP Ref_Diehl_STDPmnist_FNCOM15 under different timesteps: (a) accuracy profiles; (b) latency and energy consumption normalized to timestep 350.
• Figure 3: An overview of our novel contributions (shown in blue boxes).
• Figure 4: The neuronal dynamics of a LIF neuron model.
• Figure 5: The overview of our TopSpark methodology, where the novel steps are highlighted in blue boxes. We first analyze the impact of timestep reduction on the accuracy profiles (Section III-A). We also identify the roles of neuron parameters in different timesteps (Section III-B). Then, we leverage previous observations to enhance the neuron parameters (Section III-C) as well as develop a strategy to trade-off accuracy, latency, and energy (Section III-D). Output of the TopSpark methodology is an optimized SNN model with optimized timestep, which is employed on autonomous mobile agents. Furthermore, the mobile agents can adjust the timestep of the SNN processing at run time to adaptively meet the power/energy requirements (e.g., to save battery life).