Research map / Architectures and efficiency / Spiking networks
Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation
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.
Proposes a parallel ANN-SNN conversion framework that maps QCFS/ReLU activations onto parallel spiking neurons with optimal shift terms, distribution-aware error calibration, and binary-search spike generation for fast, low-latency SNN inference.
Problem
ANN-SNN conversion yields SNNs needing many time-steps for ANN-level accuracy, and serial LIF inference is slow; STBP training is memory- and time-expensive.
Approach
Use parallel spiking neurons with a time-step-wise shift term and pre/post conversion matrices so that the spike count over T steps exactly reproduces a QCFS quantized activation (lossless in expectation), extended by DA-QCFS calibration to ReLU networks in a training-free setting.
Key results
- Achieves 72.90% top-1 on ImageNet-1k with ResNet-34 at T=4 using error calibration (vs. COS 72.66% at T=10).
- Achieves 73.92% on ImageNet-1k VGG-16 at T=8, exceeding COS (T=10) by 3.33%.
- Training-free ResNet-18/34 reach 66.26%/68.04% at T=16, and accuracy loss within about 1% at T=32; parallel inference yields 19-38x speedup over serial IF for T>=32.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add parallel conversion to QCFS conversion | VGG-16 · ImageNet-1k, T=8 | top-1 accuracy: 73.92% ... exceeds the performance of COS (T=10) by 3.33% | Helped |
| add layer-wise error calibration (ablation) | Parallel Conversion on QCFS VGG-16 · ImageNet-1k, T=4 | top-1 accuracy: 71.23 -> 71.75 | Helped |
| add layer-wise error calibration (ablation) | Parallel Conversion on QCFS ResNet-34 · ImageNet-1k, T=4 | top-1 accuracy: 67.28 -> 72.90 | Helped |
| use training-free parallel conversion | ReLU ResNet-18 · ImageNet-1k, T=16 versus TBC at T=32 | top-1 accuracy: 66.26% ... exceeds ... TBC within 32 steps by 15.61% | Helped |
| use training-free parallel conversion | ReLU ResNet-34 · ImageNet-1k, T=16 versus TBC at T=32 | top-1 accuracy: 68.04% ... exceeds ... TBC within 32 steps by 9.01% | Helped |
| use parallel rather than serial inference | converted SNNs · ImageNet-1k, T≥32 | inference speedup: 19∼ 38× | Helped |
Surprising findings
- Parallel conversion with the optimal shift term can be lossless in expectation even when simulation latency differs from inference latency.
Limitations
- Accuracy loss remains at very small T for training-free conversion (e.g., -14.58% for ResNet-18 at T=8)
- Theoretical lossless property relies on uniform input-distribution assumptions for the simulation-latency mismatch case
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-1k
Benchmarks: CIFAR-10 CIFAR-100 ImageNet-1k
Metrics: top-1 accuracy inference speed (speedup ratio)
Baselines: STBP training (Dspike, RecDis, GLIF, TAB) OPT QCFS SNM Burst COS SRP SNNC TBC
Builds on: QCFS parallel spiking neuron ClipReLU SNNC TBC
Closest papers by meaning
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.93)
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.93)
- Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons (ICML 2025 · similarity 0.92)
- Efficient ANN-SNN Conversion with Error Compensation Learning (ICML 2025 · similarity 0.91)
- Efficient Parallel Training Methods for Spiking Neural Networks with Constant Time Complexity (ICML 2025 · similarity 0.90)
- Efficient ANN-Guided Distillation: Aligning Rate-based Features of Spiking Neural Networks through Hybrid Block-wise Replacement (CVPR 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.