Research map / Architectures and efficiency / Spiking networks
Efficient ANN-SNN Conversion with Error Compensation Learning
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 an ANN-to-SNN conversion framework combining a learnable-threshold clipping activation, dual-threshold neurons, and optimal membrane potential initialization to reach high accuracy at ultra-low latency (2-4 time steps).
Problem
ANN-to-SNN conversion suffers large accuracy loss at low latency due to clipping, quantization, and uneven-activation conversion errors, requiring many time steps.
Approach
Jointly mitigates the three conversion errors: a learnable-threshold clip-floor-shift activation trained in the ANN so its threshold maps directly to the SNN threshold (removing clipping error); dual-threshold neurons that emit negative spikes to cancel over-fired spikes (reducing quantization and unevenness error); and initializing membrane potential to half the threshold.
Key results
- Achieves 94.75% accuracy on CIFAR-10 with ResNet-18 at T=2 (vs. 75.44% for QCFS-converted SNN at T=2 per the paper's table).
- On ImageNet with VGG-16 at T=16, reaches 74.15% accuracy, 23.18% higher than QCFS.
- On CIFAR-100 VGG-16, energy at T=4 is 0.008 mJ vs 4.170 mJ for the source ANN, with 76.50% accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a dual-threshold neuron (ablation) | converted ResNet-18 SNN · CIFAR-10, T=2 | accuracy: 75.44% -> 94.75%; +19.31% | Helped |
| Add a dual-threshold neuron (ablation) | converted ResNet-20 SNN · CIFAR-100, T=2 | accuracy: 19.96% -> 42.91%; +22.95% | Helped |
| Increase quantization steps from L=2 to L=4 (ablation) | converted VGG-16 SNN · CIFAR-10, T=2 | accuracy: 94.35% -> 93.53% | Hurt |
| Increase quantization steps from L=2 to L=4 (ablation) | converted VGG-16 SNN · CIFAR-100, T=2 | accuracy: 73.78% -> 72.58% | Hurt |
| Increase quantization steps from L=2 to L=4 (ablation) | converted VGG-16 SNN · CIFAR-10, T=4 | accuracy: 95.04% -> 95.15% | Mixed |
| Increase quantization steps from L=2 to L=4 (ablation) | converted VGG-16 SNN · CIFAR-100, T=4 | accuracy: 75.70% -> 76.50% | Helped |
| Use the proposed conversion instead of QCFS | ResNet-18 SNN · CIFAR-10, T=2 | accuracy: 75.44% -> 94.75% | Helped |
| Use the proposed conversion instead of QCFS | ResNet-34 SNN · ImageNet, T=16 | accuracy: 59.35% -> 72.37% | Helped |
| Use the proposed conversion instead of QCFS | VGG-16 SNN · CIFAR-100, T=2 | accuracy: 63.79% -> 72.58% | Helped |
| Use the proposed conversion at T=16 instead of ANN inference | VGG-16 · CIFAR-100 | energy: 4.170 -> 0.034 | Helped |
Surprising findings
- Reported ResNet-20 CIFAR-100 ANN accuracy (68.58%) for the proposed method is below the QCFS ANN baseline (69.94%) in the main comparison table.
Limitations
- Energy analysis is only on CIFAR-100 with VGG-16 and uses per-operation energy estimates rather than hardware measurements.
- Some ImageNet table entries are missing (e.g., T=256 for proposed method).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet
Benchmarks: CIFAR-10 CIFAR-100 ImageNet
Metrics: top-1 accuracy energy consumption (mJ)
Baselines: RMP TSC RTS RNL Opt QCFS SNNC-AP
Builds on: QCFS Opt SRP RMP
Closest papers by meaning
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.93)
- Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation (ICML 2025 · similarity 0.91)
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.91)
- Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons (ICML 2025 · similarity 0.90)
- Differential Coding for Training-Free ANN-to-SNN Conversion (ICML 2025 · similarity 0.90)
- LM-HT SNN: Enhancing the Performance of SNN to ANN Counterpart through Learnable Multi-hierarchical Threshold Model (NeurIPS 2024 · similarity 0.88)
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.