Research map / Architectures and efficiency / Spiking networks
Differential Coding for Training-Free ANN-to-SNN Conversion
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 differential coding (DCGS), a training-free ANN-to-SNN conversion scheme that transmits changes in encoded activation rather than rates, plus a threshold iteration method for ReLU thresholds, evaluated on CNNs and Transformers on ImageNet.
Problem
Rate-coding-based ANN-to-SNN conversion needs many spikes and long time-steps, raising energy and latency; conversion of nonlinear layers and Transformers is also hard without retraining.
Approach
Differential coding treats time-weighted spikes as corrections to the encoded activation (updated only when spikes occur) instead of averaging firing rates, avoiding the decay of rate coding; it is combined with differential graded units for nonlinear/two-input ops and a threshold iteration method that minimizes expected quantization and clipping error under a Gaussian activation assumption.
Key results
- On ImageNet, DCGS converts VGG-16 (73.25% ANN) to 73.17% SNN accuracy at T=8 with energy ratio 0.22, per the reported table.
- ResNet34-4/1 reaches 76.04% at T=8 (ANN 76.42%) with energy ratio 0.37.
- ViT-S-8/4 reaches 81.11% at T=4 (ANN 81.38%); accuracy and energy ratio grow with T for Transformers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use differential coding instead of rate coding | ANN-to-SNN conversion · ImageNet | accuracy | Helped |
| Use differential coding instead of rate coding | ANN-to-SNN conversion · ImageNet | energy consumption | Helped |
| Use threshold iteration instead of the 99.9% large activation method | ResNet34 ANN-to-SNN conversion · ImageNet | accuracy | Helped |
| Use threshold iteration instead of the 99.9% large activation method | ResNet34 ANN-to-SNN conversion · ImageNet | energy consumption | Helped |
| Increase time-steps from T=2 to T=8 | DCGS with ResNet34-4/1 · ImageNet | accuracy: 59.71 -> 76.04 | Helped |
| Increase time-steps from T=2 to T=16 | DCGS with ViT-Small-8/4 · ImageNet | accuracy: 77.84 -> 81.38 | Helped |
Surprising findings
- Transformer conversion energy ratio rises with time-steps (e.g., ViT-S 0.32 at T=2 to 1.71 at T=16), in contrast to CNNs where accuracy gains come with modest energy growth.
Limitations
- Requires at least one negative threshold to generate negative spikes for error correction; otherwise spike errors accumulate.
- No optimal threshold method for Transformers; threshold iteration not suitable for them, which limits Transformer performance.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet
Metrics: top-1 accuracy energy ratio
Baselines: TS SNM MMSE QCFS SRP MST STA SpikeZIP-TF ECMT
Builds on: QCFS multi-threshold neurons rate coding ANN-to-SNN conversion
Closest papers by meaning
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.90)
- Efficient ANN-SNN Conversion with Error Compensation Learning (ICML 2025 · similarity 0.90)
- Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation (ICML 2025 · similarity 0.89)
- TTFSFormer: A TTFS-based Lossless Conversion of Spiking Transformer (ICML 2025 · similarity 0.89)
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · similarity 0.88)
- Efficient ANN-Guided Distillation: Aligning Rate-based Features of Spiking Neural Networks through Hybrid Block-wise Replacement (CVPR 2025 · 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.