Research map / Label-efficient and robust learning / Domain adaptation and generalization
An Efficient Knowledge Transfer Strategy for Spiking Neural Networks from Static to Event Domain
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.
Transfers knowledge from static images to event-based SNN training via a CKA-based domain alignment loss with learnable spatio-temporal regularization and a probabilistic sliding training schedule that gradually swaps static inputs for event data.
Problem
SNNs overfit on small event-based (DVS) datasets, and naively pre-training on static images then fine-tuning on event data transfers poorly due to a domain mismatch in feature distributions.
Approach
A knowledge transfer loss combining CKA-based domain alignment between static-image and event features per time step, weighted by learnable sigmoid coefficients that trade off against the event classification loss (TET), plus a sliding schedule that probabilistically replaces static inputs with event data over training.
Key results
- On N-Caltech101 with VGGSNN (T=10), the method reaches 93.18 ± 0.38 (best 93.45) vs 85.53 for ETC and 79.27 for TET.
- On CEP-DVS with ResNet-18, accuracy rises from 25.05 (TET, reproduced) to 30.05 ± 0.50.
- On N-Omniglot with SCNN, accuracy is 63.60 ± 0.46 vs 63.00 for the plain baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add knowledge transfer to VGGSNN | VGGSNN · N-Caltech101 | accuracy: 85.53 ± 0.09 -> 93.18 ± 0.38 (93.45) | Helped |
| Add knowledge transfer to ResNet-18 | ResNet-18 · CEP-DVS | accuracy: 25.05 ± 0.66 (25.70)* -> 30.05 ± 0.50 (30.50) | Helped |
| Add knowledge transfer to SCNN | SCNN · N-Omniglot | accuracy: 63.00 ± 0.32 (63.44)* -> 63.60 ± 0.46 (64.09) | Helped |
| Add DAL to the baseline (ablation) | VGGSNN · N-Caltech101 | accuracy: 79.66% -> 82.07% | Helped |
| Add STR to DAL without sliding training (ablation) | VGGSNN · N-Caltech101 | accuracy: 82.07% -> 83.56% | Helped |
| Add sliding training to DAL and STR (ablation) | VGGSNN · N-Caltech101 | accuracy: 83.56% -> 92.64% | Helped |
| Add DAL and STR to the no-KTL variant (ablation) | VGGSNN · N-Caltech101 | accuracy: 84.14% -> 92.64% | Helped |
| Add DAL and STR to the no-KTL variant (ablation) | ResNet-18 · CEP-DVS | accuracy: 27.55% -> 30.50% | Helped |
| Add sliding training (ablation) | VGGSNN · N-Caltech101 | accuracy: 83.56% -> 92.64% | Helped |
| Add sliding training (ablation) | ResNet-18 · CEP-DVS | accuracy: 23.70% -> 30.50% | Helped |
| Replace CKA with MMD (ablation) | ResNet-18 with knowledge transfer · CEP-DVS | accuracy: 30.50% -> 26.25% | Hurt |
Surprising findings
- Naive static-pretrain then event-finetune, or using the method without sliding training, can underperform direct training on CEP-DVS (23.70% vs 25.70%).
- Replacing CKA with MMD as the alignment metric gives only a small gain (26.25% vs 30.50% for CKA on CEP-DVS).
Limitations
- Gains on N-Omniglot are small, attributed to only about 20 static images per class.
- Requires paired static images of the same class for each event sample and extra training computation.
- Sensitive to training schedule: without sliding training, performance on CEP-DVS can drop below baseline.
Experimental setup
Datasets: N-Caltech101 Caltech101 CEP-DVS CIFAR100 N-Omniglot Omniglot
Benchmarks: N-Caltech101 CEP-DVS N-Omniglot
Metrics: classification accuracy
Baselines: NDA EventMixer TET TJCA-TET TKS ETC SEW-ResNet18 Spikformer
Builds on: TET BrainCog
Closest papers by meaning
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.88)
- Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks (CVPR 2025 · similarity 0.87)
- Efficient Logit-based Knowledge Distillation of Deep Spiking Neural Networks for Full-Range Timestep Deployment (ICML 2025 · similarity 0.87)
- Advancing Spiking Neural Networks Towards Multiscale Spatiotemporal Interaction Learning (AAAI 2025 · similarity 0.87)
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · similarity 0.87)
- Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons (ICML 2025 · similarity 0.87)
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.