Energy-Aware Spike Budgeting for Continual Learning in SNNs: Our NCE 2026 Paper

I am proud to share that our paper, “Energy-aware spike budgeting for continual learning in spiking neural networks for neuromorphic vision”, has been published in Neuromorphic Computing and Engineering (IOP Publishing).

This work is especially meaningful to me as a mentor. The first authors are my former students at the University of Liberal Arts Bangladesh (ULAB)—Anika Tabassum Meem and Muntasir Hossain Nadid—whom I supervised while serving as Lecturer in the Department of Electrical and Electronic Engineering. I am the corresponding author on the paper, continuing that mentorship from my ULAB years into a full journal publication.

The Problem: Continual Learning Under Spike Cost

Neuromorphic vision systems based on spiking neural networks (SNNs) promise event-driven, sparse computation for both frame-based and event-based cameras. Yet catastrophic forgetting—the abrupt loss of previously learned knowledge when new tasks arrive—remains a central barrier to deployment in continually changing environments.

Most continual learning methods were developed for conventional artificial neural networks. They rarely jointly optimize task accuracy and activity-dependent spike cost, and exploration on event-based datasets has been particularly limited. Fixed spike-rate penalties also struggle across modalities: a coefficient that sparsifies dense frame encodings can over-constrain sparse DVS streams, while one tuned for event data may not sufficiently regularize frame inputs as the replay distribution evolves.

Our Approach: Adaptive Spike Budgeting

We propose an energy-aware spike budgeting framework for continual SNN learning that integrates:

1. Experience replay to mitigate forgetting across a task stream
2. Learnable leaky integrate-and-fire (LIF) neuron parameters so dynamics can adapt during training
3. An adaptive spike-budget controller that enforces dataset-specific spike-activity constraints

Rather than a static sparsity weight, the controller closes a feedback loop by comparing observed mini-batch activity with a target budget during replay. The same control law can tighten or relax the activity constraint as the task stream and input modality change.

Modality-Dependent Duality

A central finding is an operational duality in SNN continual learning:

• 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, the method improves the accuracy–spike-activity trade-off while keeping spike activity under explicit control—using an activity-driven synaptic-event proxy rather than a direct hardware-energy measurement.

Why Mentorship Matters

Beyond the technical contributions, this project reflects what academic mentorship can unlock: undergraduate researchers at ULAB driving a full research stack—from problem formulation through experiments to a peer-reviewed journal paper. Supervising this work reinforced my commitment to open, collaborative science and to creating pathways for students into neuromorphic and ML research.

Read the Paper

• DOI (NCE)
• arXiv:2602.12236
• PDF

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