Research map / Architectures and efficiency / Spiking networks
Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks
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 AdaFire, SSC and IAT, three training-free add-ons to calibration-based ANN-to-SNN conversion that tune per-layer burst firing, spike thresholds and per-input timesteps to cut latency and energy.
Problem
ANN-to-SNN conversion via calibration is training-free but fails to reach high accuracy within few timesteps, and converted SNNs' energy efficiency is not ensured.
Approach
Extends spike calibration with a burst-spike neuron (up to φ spikes per timestep) whose per-layer φ is chosen by a KL-based layer sensitivity and Pareto/dynamic-programming search under an energy budget; adds sensitivity-driven threshold amplification (spike compression) and entropy-based input-aware early exit with timestep-dependent thresholds.
Key results
- On ImageNet VGG-16 at T=8, AdaFire reaches 73.53% vs 25.33% for Calibration at T=8 (ANN 75.36%).
- The full framework (AdaFire+SSC+IAT) cuts CIFAR-10 energy by 70.12% (14.86 to 4.44 mJ) with accuracy 95.47% vs 96.34% baseline.
- On CIFAR10-DVS, AdaFire reports 81.25% accuracy at 8 timesteps versus 80.90% for Spikformer at 10 timesteps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Adaptive-Firing Neuron Model (AdaFire) added to calibration (ablation) | SNN calibration · ImageNet ablation | Accuracy (%): 56.74 -> 68.45 (+11.71) | Helped |
| Adaptive-Firing Neuron Model (AdaFire) added to calibration (ablation) | SNN calibration · CIFAR-10 ablation | Accuracy (%): 96.34 -> 96.69 (+0.35) | Helped |
| Adaptive Sensitivity Spike Compression (SSC) added (ablation) | AdaFire-calibrated SNN · CIFAR-10 ablation | Energy (mJ): 20.49 -> 7.71 (-48.12%) | Helped |
| Input-aware Adaptive Timesteps (IAT) added (ablation) | AdaFire-calibrated SNN · CIFAR-10 ablation | Energy (mJ): 20.49 -> 7.06 (-52.48%) | Helped |
| Full combination of AdaFire, SSC and IAT (ablation) | SNN calibration · ImageNet ablation | Energy (mJ): 162.56 -> 92.50 (-43.10%) | Helped |
| Full combination of AdaFire, SSC and IAT (ablation) | SNN calibration · CIFAR-10 ablation | Accuracy (%): 96.34 -> 95.47 (-0.87) | Mixed |
| Adaptive-Firing (AdaFire) with per-layer firing patterns | VGG-16 calibration · ImageNet T=8 | Top-1 accuracy (%): 25.33 -> 73.53 | Helped |
| AdaFire on object detection | YOLOv2 calibration · COCO 2017 mAP | mAP: 21.79 -> 28.04 (+6.25) with 16 vs 128 timesteps | Helped |
| SSC applied to ImageNet | SNN calibration · ImageNet | Energy saving: 32.4% energy saving with 0.5% accuracy drop | Helped |
Surprising findings
- The full combined framework on CIFAR-10 drops accuracy to 95.47% from 96.34% baseline, even though AdaFire alone raised it to 96.69%.
Limitations
- Assumes per-layer sensitivities are independent, simplifying the joint search.
- Energy is estimated from spike counts times a per-spike cost rather than measured on hardware.
- Paper claims 'training-free' but the appendix details are not in the provided text and no code release is stated.
Experimental setup
Datasets: ImageNet CIFAR-10 CIFAR-100 CIFAR10-DVS N-Caltech101 N-Cars ShapeNet PASCAL VOC 2012 MS COCO 2017
Benchmarks: ImageNet CIFAR10-DVS N-Caltech101 N-Cars PASCAL VOC 2012 MS COCO 2017 ShapeNet CIFAR-10 CIFAR-100
Metrics: top-1 accuracy mAP energy consumption (mJ) timesteps
Baselines: OPT SNM QCFS Calibration TA-SNN PLIF Dspkie DSR Spikformer SALT NDA CarSNN
Builds on: spike calibration QCFS Integrate-and-Fire neuron PLIF
Closest papers by meaning
- Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation (ICML 2025 · similarity 0.93)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.92)
- Efficient ANN-SNN Conversion with Error Compensation Learning (ICML 2025 · similarity 0.91)
- Training High Performance Spiking Neural Network by Temporal Model Calibration (ICML 2025 · similarity 0.90)
- Rethinking the Membrane Dynamics and Optimization Objectives of Spiking Neural Networks (NeurIPS 2024 · similarity 0.90)
- Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons (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.