Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Energy-Aware Spike Budgeting for Continual Learning in Spiking Neural Networks for Neuromorphic Vision

Anika Tabassum Meem, Muntasir Hossain Nadid, Md Zesun Ahmed Mia

TL;DR

Neuromorphic vision with spiking neural networks faces catastrophic forgetting under continual learning while energy constraints limit hardware viability. The authors propose an energy-aware spike budgeting framework that jointly optimizes accuracy and energy by combining experience replay, learnable LIF neuron dynamics, and an adaptive spike scheduler. Across five benchmarks spanning frame- and event-based vision, the approach delivers consistent accuracy gains and substantial spike reductions on frame data, while enabling targeted spike increases for event streams to unlock large performance improvements with minimal energy cost. The results demonstrate modality-dependent benefits and establish a practical pathway for deploying continual learning in neuromorphic hardware, with clear avenues for automatic budget tuning and on-hardware validation.

Abstract

Neuromorphic vision systems based on spiking neural networks (SNNs) offer ultra-low-power perception for event-based and frame-based cameras, yet catastrophic forgetting remains a critical barrier to deployment in continually evolving environments. Existing continual learning methods, developed primarily for artificial neural networks, seldom jointly optimize accuracy and energy efficiency, with particularly limited exploration on event-based datasets. We propose an energy-aware spike budgeting framework for continual SNN learning that integrates experience replay, learnable leaky integrate-and-fire neuron parameters, and an adaptive spike scheduler to enforce dataset-specific energy constraints during training. Our approach exhibits modality-dependent behavior: on frame-based datasets (MNIST, CIFAR-10), spike budgeting acts as a sparsity-inducing regularizer, improving accuracy while reducing spike rates by up to 47\%; on event-based datasets (DVS-Gesture, N-MNIST, CIFAR-10-DVS), controlled budget relaxation enables accuracy gains up to 17.45 percentage points with minimal computational overhead. Across five benchmarks spanning both modalities, our method demonstrates consistent performance improvements while minimizing dynamic power consumption, advancing the practical viability of continual learning in neuromorphic vision systems.

Energy-Aware Spike Budgeting for Continual Learning in Spiking Neural Networks for Neuromorphic Vision

TL;DR

Neuromorphic vision with spiking neural networks faces catastrophic forgetting under continual learning while energy constraints limit hardware viability. The authors propose an energy-aware spike budgeting framework that jointly optimizes accuracy and energy by combining experience replay, learnable LIF neuron dynamics, and an adaptive spike scheduler. Across five benchmarks spanning frame- and event-based vision, the approach delivers consistent accuracy gains and substantial spike reductions on frame data, while enabling targeted spike increases for event streams to unlock large performance improvements with minimal energy cost. The results demonstrate modality-dependent benefits and establish a practical pathway for deploying continual learning in neuromorphic hardware, with clear avenues for automatic budget tuning and on-hardware validation.

Abstract

Neuromorphic vision systems based on spiking neural networks (SNNs) offer ultra-low-power perception for event-based and frame-based cameras, yet catastrophic forgetting remains a critical barrier to deployment in continually evolving environments. Existing continual learning methods, developed primarily for artificial neural networks, seldom jointly optimize accuracy and energy efficiency, with particularly limited exploration on event-based datasets. We propose an energy-aware spike budgeting framework for continual SNN learning that integrates experience replay, learnable leaky integrate-and-fire neuron parameters, and an adaptive spike scheduler to enforce dataset-specific energy constraints during training. Our approach exhibits modality-dependent behavior: on frame-based datasets (MNIST, CIFAR-10), spike budgeting acts as a sparsity-inducing regularizer, improving accuracy while reducing spike rates by up to 47\%; on event-based datasets (DVS-Gesture, N-MNIST, CIFAR-10-DVS), controlled budget relaxation enables accuracy gains up to 17.45 percentage points with minimal computational overhead. Across five benchmarks spanning both modalities, our method demonstrates consistent performance improvements while minimizing dynamic power consumption, advancing the practical viability of continual learning in neuromorphic vision systems.
Paper Structure (41 sections, 12 equations, 2 figures, 15 tables)

This paper contains 41 sections, 12 equations, 2 figures, 15 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. LIF Neuron Model and Surrogate Gradients
  4. Spike-Rate as an Energy Proxy
  5. Continual Learning Metrics
  6. Method: Energy-Aware Continual Learning in SNNs
  7. Spiking Network Architectures
  8. Energy-Aware Spike Budgeting Framework
  9. Learnable Neuron Dynamics
  10. Continual Learning Components
  11. Class-Incremental Protocol
  12. Experience Replay
  13. Ablation Framework
  14. Experiments
  15. Experimental Setup
  16. ...and 26 more sections

Figures (2)

  • Figure 1: Accuracy-energy trade-off across representative datasets. Arrows indicate improvement from C1 (hollow circles) to best configuration (filled circles). MNIST achieves both accuracy improvement and spike reduction. DVS datasets show controlled spike increases that enable substantial accuracy gains while remaining ultra-sparse (below $1\%$ on DVS-Gesture and below $8\%$ across all DVS benchmarks).
  • Figure 2: Task-wise analysis for representative datasets. (a,b) DVS-Gesture: C4 maintains ultra-low spike rates below 1% while improving accuracy. (c,d) CIFAR-10: C4 progressively reduces spike rates across tasks.