Temporal Misalignment in ANN-SNN Conversion and Its Mitigation via Probabilistic Spiking Neurons

TL;DR

The paper identifies temporal misalignment as a spike-timing effect in ANN-SNN conversion that can unexpectedly boost low-latency performance. It introduces two-phase probabilistic spiking (TPP) neurons that accumulate inputs before probabilistic firing to emulate spike-permutation effects and closely approximate ReLU activations, with theoretical and practical underpinnings. Empirical results show state-of-the-art or competitive accuracy across CIFAR-10/100, CIFAR10-DVS, and ImageNet, including strong gains when combined with QCFS or SNNC baselines, and evidence of hardware-friendly feasibility. Overall, the work advances energy-efficient SNN deployment by improving conversion quality and providing a biologically plausible, hardware-mappable neuronal mechanism.

Abstract

Spiking Neural Networks (SNNs) offer a more energy-efficient alternative to Artificial Neural Networks (ANNs) by mimicking biological neural principles, establishing them as a promising approach to mitigate the increasing energy demands of large-scale neural models. However, fully harnessing the capabilities of SNNs remains challenging due to their discrete signal processing and temporal dynamics. ANN-SNN conversion has emerged as a practical approach, enabling SNNs to achieve competitive performance on complex machine learning tasks. In this work, we identify a phenomenon in the ANN-SNN conversion framework, termed temporal misalignment, in which random spike rearrangement across SNN layers leads to performance improvements. Based on this observation, we introduce biologically plausible two-phase probabilistic (TPP) spiking neurons, further enhancing the conversion process. We demonstrate the advantages of our proposed method both theoretically and empirically through comprehensive experiments on CIFAR-10/100, CIFAR10-DVS, and ImageNet across a variety of architectures, achieving state-of-the-art results.

Figures (14)

• Figure 1: The initial experiment: After ANN-SNN conversion, we compared the accuracy of the baseline model QCFS bu2022optimal with its "permuted" version and our proposed TPP neurons (setting is VGG16 - ImageNet, ANN acc. 74.29%).
• Figure 2: Spike train permutation: spikes at different time steps are shuffled to alter their temporal order.
• Figure 3: Two-Phase Probabilistic (TPP) spiking neuron mechanism operates in two phases: accumulation of inputs and probabilistic spiking based on membrane potential. The model uses a Bernoulli process for spike generation, with membrane potential updates over time steps.
• Figure 4: The membrane potential distributions of the first channel (randomly selected) across three modes (baseline, shuffle, and probabilistic) in VGG-16 on CIFAR-100. For comparative analysis, the first two timesteps ($t=1$, $t=2$) from a total of eight timesteps ($T=8$) are selected for each mode. The baseline mode (blue) attains an accuracy of 24.22%, while the shuffle mode (light green) enhances accuracy to 70.54%, and the probabilistic mode (dark orange) further improves accuracy to 73.42%. The distributions are presented prior to neuronal firing, with the red dashed line indicating the threshold voltage (Vth) for the respective layer (see Appendix \ref{['appendix:mem-pot-dist']}).
• Figure 5: Spike counts of VGG-16 on CIFAR-100 of RTS baseline compared with RTS+TPP. Note: The bar height from bottom indicates the spike counts after each timestep T, and the color of longer Ts is overlaid by shorter Ts (see Appendix \ref{['appendix:firing-counts']})

Theorems & Definitions (5)

• Theorem 1
• Theorem 1
• proof
• Theorem 1
• proof