Parameter Scope Matters: MNE-L2 for Reliable Low-Latency ANN-to-SNN Conversion

Spiking neural networks (SNNs) enable event-driven inference by representing neural activations as discrete spikes, offering the potential for low-power deployment on neuromorphic hardware. One established approach to construct deep SNNs is to first train a conventional artificial neural network (ANN) and then convert its activations into spike rates or spike counts. However, this conversion makes inference sensitive to perturbations: small changes in ANN activations can cause a neuron to emit a different number of spikes. This raises an important question: how do training choices in the source ANN affect the robustness of the converted SNN? We study this question through the scope of L2 regularization. We find that the common practice of applying L2 regularization to all trainable parameters, including the affine scale parameters of batch normalization (BN), can substantially reduce the noise robustness of converted SNNs compared with regularizing weights alone. Based on this observation, we propose margin-normalized effective L2 (MNE-L2), which adapts the regularization strength of each source-ANN weight according to its BN-folded amplification and its distance to the nearest spike-count threshold after conversion. Across five-seed CIFAR experiments, MNE-L2 improves high-noise accuracy over all-parameter L2 by up to 62.06 percentage points while reducing clean firing activity and estimated arithmetic energy. These findings expose regularization scope as an important training choice and demonstrate that MNE-L2 improves the reliability-efficiency trade-off for low-latency SNN inference.