Forward-Only Test-Time Adaptation for Spiking Neural Networks

Spiking neural networks promise efficient temporal inference, but their accuracy can deteriorate when deployed inputs differ from training data. Adapting at test time could address this shift, yet most parameter-based methods require backpropagation through temporal states—a poor match for inference-oriented neuromorphic hardware. We ask whether a pretrained spiking network can instead adapt online using forward evaluations alone. Naive gradient-free updates are unreliable because small parameter perturbations can trigger discontinuous spike changes and amplify estimator noise. Our approach updates only a small channel adapter and stabilizes symmetric loss comparisons in two complementary ways: perturbing fewer coordinates or averaging multiple directions, while reusing the same sampled spike train across each comparison. Across disjoint CIFAR-10-C streams, forward-only adaptation matches the frozen model and a comparable gradient-based update, while outperforming standard adaptation baselines by up to 17 percentage points. It reduces peak memory by 17× relative to a full-graph gradient baseline and by 127× relative to an augmentation-based baseline. The result is a practical alternative when repeated inference is available but backward execution or temporal-graph storage is not.