Enhancing Spiking Neural Networks with Online Test-Time Adaptation for Distribution Shifts

Researchers from Southern University of Science and Technology, Chongqing University, and Huawei Technologies have developed an innovative method to address a critical challenge in deploying spiking neural networks (SNNs) on neuromorphic chips: their ability to adapt to distribution shifts in real-world scenarios. Their approach, called Threshold Modulation (TM), enables SNNs to dynamically adjust to new data distributions during deployment without requiring source data or labeled target samples.

SNNs have gained popularity as energy-efficient alternatives to traditional artificial neural networks (ANNs), particularly for edge devices with low power requirements. These networks transmit information through discrete spike events rather than continuous activations, mimicking biological neural systems. When deployed on specialized neuromorphic chips, SNNs offer remarkable energy efficiency for tasks like image classification. However, like all neural networks, their performance can degrade significantly when facing input distribution shifts, such as environmental changes or sensor aging.

The researchers identified a significant gap in the field: while online test-time adaptation (OTTA) techniques exist for traditional ANNs, these methods are not well-suited for SNNs, especially when deployed on neuromorphic hardware with its unique constraints. Their novel TM approach addresses this gap through neuronal dynamics-inspired normalization that modulates the firing threshold of spiking neurons.

The TM framework consists of three phases: pre-training, deployment, and online adaptation. During pre-training, the SNN learns features from the source domain using Membrane Potential Batch Normalization (MPBN), which normalizes membrane potentials after neuronal charging. Before deployment, the model undergoes threshold re-parameterization, allowing it to be mapped onto neuromorphic hardware. During adaptation, the model dynamically adjusts firing thresholds based on current statistics, making it responsive to distribution shifts without modifying model weights or requiring additional computational overhead.

The researchers developed two main variants of their approach: TM-NORM, which keeps affine parameters frozen while updating statistics, and TM-ENT, which additionally updates affine parameters through entropy minimization. Extensive experiments on benchmark datasets including CIFAR-10/100-C, ImageNet-C, and various digit recognition transfer tasks demonstrated the effectiveness of both variants.

For example, on CIFAR-10-C with the ResNet-20 backbone, TM-NORM reduced the average classification error from 46.7% to 25.2%, while TM-ENT achieved a slightly better 24.8%. Similar improvements were observed across different model architectures and datasets. The method proved effective even in resource-constrained settings with a batch size of 1, showing its practical applicability for edge devices.

Energy consumption analysis revealed that TM-NORM increases power consumption by only 3% compared to models without adaptation capability, making it significantly more efficient than direct calibration approaches that require 12% additional energy. The researchers also conducted ablation studies to understand the contribution of different components of their approach, finding that entropy minimization offers limited benefits while substantially increasing energy consumption.

This work represents one of the first dedicated attempts to address online test-time adaptation for SNNs in neuromorphic computing environments. By enabling SNNs to adapt to changing data distributions with minimal computational overhead, the proposed method enhances their practical utility in real-world deployment scenarios and provides valuable guidance for future neuromorphic chip design.