Research map / Architectures and efficiency / Spiking networks
Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Identifies "temporal misalignment" in ANN-SNN conversion (randomly permuting spike trains after spiking layers improves accuracy at low latency) and proposes two-phase probabilistic (TPP) spiking neurons that emulate this effect, combined with QCFS, RTS, and SNNC conversion methods.
Problem
ANN-SNN conversion SNNs perform poorly at low latency because spike timing is suboptimal; the paper asks how spike temporal arrangement affects accuracy and how to build neurons that exploit it.
Approach
Random permutation of spike trains after each spiking layer improves converted-SNN accuracy; the authors design two-phase neurons that accumulate input and then fire via Bernoulli sampling with probability v/(θ(T-t+1)), which provably matches ReLU clipping in expectation and retains the stochasticity of permutation.
Key results
- On ImageNet VGG-16 at T=128, TPP+QCFS reaches 74.33% vs 74.29% ANN accuracy; at T=32, TPP+QCFS gets 74.23%.
- On ImageNet ResNet-34 with QCFS at T=16, TPP improves accuracy from 59.35% to 72.03%.
- On CIFAR10-DVS, TPP+SNNC reaches 82.40% at 8 timesteps, above AdaFire (81.25%) and Spikformer (80.90%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace QCFS spiking neurons with TPP neurons | QCFS on ImageNet ResNet-34 · ImageNet, ResNet-34, T=16 | accuracy: 59.35% to 72.03% | Helped |
| Replace QCFS spiking neurons with TPP neurons | QCFS on ImageNet VGG-16 · ImageNet, VGG-16, T=16 | accuracy: 50.97% to 73.98% | Helped |
| Use TPP neurons instead of the SNNC-AP conversion | SNNC conversion on ImageNet VGG-16 · ImageNet, VGG-16, T=32 | accuracy: 63.64 -> 74.72 (0.06) | Helped |
| Use TPP neurons instead of the SNNC-AP conversion | SNNC conversion on CIFAR-100 VGG-16 · CIFAR-100, VGG-16, T=8 | accuracy: 64.30 -> 73.16 (0.17) | Helped |
| Permute spike trains after spiking layers | ANN-SNN conversion · VGG-16, CIFAR-100, T=8 | accuracy: 24.22% to 70.54% | Helped |
| Use probabilistic TPP spiking instead of the baseline neuron | VGG-16 on CIFAR-100 · CIFAR-100, VGG-16, T=8 | accuracy: 24.22% to 73.42% | Helped |
| Use TPP instead of spike-train permutation | ANN-SNN conversion on CIFAR-10 · CIFAR-10, ResNet-20, T=4 | accuracy: 90.54 -> 79.19 (0.67) | Hurt |
| Use TPP instead of spike-train permutation | ANN-SNN conversion on CIFAR-10 · CIFAR-10, VGG-16, T=8 | accuracy: 92.24 -> 94.34 (0.12) | Helped |
| Use TPP instead of spike-train permutation | ANN-SNN conversion on CIFAR-100 · CIFAR-100, VGG-16, T=8 | accuracy: 69.42 -> 73.16 (0.17) | Helped |
| Use TPP instead of spike-train permutation | ANN-SNN conversion on ImageNet · ImageNet, VGG-16, T=4 | accuracy: 38.61 -> 54.14 (0.59); 74.34 -> 74.72 (0.06) | Mixed |
| Use TPP instead of spike-train permutation | ANN-SNN conversion on CIFAR-100 · CIFAR-100, ResNet-20, T=32 | accuracy: 81.12 -> 81.06 (0.05) | No clear effect |
Surprising findings
- Randomly permuting spike trains after each spiking layer improves accuracy of converted SNNs over the baseline, especially at low latency.
- TPP+SNNC without calibration reaches 82.40% on CIFAR10-DVS at T=8 while keeping spike counts roughly similar to baseline.
Limitations
- Advantage is largest at low latency; at high latency gains are marginal (e.g., +0.04 to +0.1 points).
- TPP+SNNC at T=1-2 on CIFAR-100 is much worse than SNNC-AP baseline, and some settings underperform permutation.
- Energy claims rely on spike counts rather than hardware measurements; per-sample latency may increase.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CIFAR10-DVS ImageNet
Benchmarks: CIFAR-10 CIFAR-100 CIFAR10-DVS ImageNet
Metrics: top-1 accuracy standard deviation over 5 runs firing/spike counts
Baselines: RTS SNNC-AP QCFS SNM Burst OPI SRP DDI FTBC Spikformer AdaFire
Builds on: QCFS RTS SNNC
Closest papers by meaning
- Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation (ICML 2025 · similarity 0.92)
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.91)
- TS-SNN: Temporal Shift Module for Spiking Neural Networks (ICML 2025 · similarity 0.91)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.91)
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.90)
- Training High Performance Spiking Neural Network by Temporal Model Calibration (ICML 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.