Research map / Label-efficient and robust learning / Self-supervised and contrastive
Towards More Discriminative Feature Learning in SNNs with Temporal-Self-Erasing Supervision
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 Temporal-Self-Erasing (TSE) supervision for directly trained SNNs, which erases highly activated regions using earlier time-step predictions so each time step gets distinct supervision, improving CIFAR-100, ImageNet and DVS-CIFAR10 accuracy.
Problem
Directly trained SNNs learn redundant, temporally invariant features because all time steps share one loss on averaged predictions, which limits feature discrimination across time steps.
Approach
At each time step, build an erasing mask from the averaged class-probability map of previous time steps (fixed plus dynamic threshold) and apply it to the current feature map before classification, so each time step is supervised on less-activated regions; no extra modules or inference cost.
Key results
- Achieves 81.47% accuracy on CIFAR-100 with ResNet-19 at 4 time steps, 6.80% above SLTT.
- Reaches 83.60% on DVS-CIFAR10 with ResNet-19 at 10 time steps, versus 77.90% for LSG.
- On ImageNet ResNet-34 at 4 time steps, reaches 71.04%, 0.50 points above PSN and well above MPBN's 64.71%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TSE supervision (ablation) | ResNet-19 · CIFAR-100 | accuracy: 78.01 -> 81.47 | Helped |
| Use TET supervision instead of TSE (ablation) | ResNet-19 · CIFAR-100 | accuracy: 78.01 -> 80.04 | Helped |
| Add TSE supervision instead of TET (ablation) | ResNet-19 · CIFAR-100 | accuracy: 80.04 -> 81.47 | Helped |
| Add TSE supervision (ablation) | ResNet-19 · DVS-CIFAR10 | accuracy: 79.20 -> 83.60 | Helped |
| Use TET supervision instead of the baseline (ablation) | ResNet-19 · DVS-CIFAR10 | accuracy: 79.20 -> 77.60 | Hurt |
| Use fixed-threshold mask construction (ablation) | ResNet-19 with TSE · CIFAR-100 | accuracy: 78.01 -> 80.80 | Helped |
| Use dynamic-threshold mask construction (ablation) | ResNet-19 with TSE · CIFAR-100 | accuracy: 78.01 -> 77.19 | Hurt |
| Combine fixed and dynamic thresholds for mask construction (ablation) | ResNet-19 with TSE · CIFAR-100 | accuracy: 78.01 -> 81.47 | Helped |
| Set κ=0 and τf=0.6 for hybrid thresholding (ablation) | ResNet-19 with TSE · CIFAR-100 | accuracy: 81.47 | Helped |
Surprising findings
- Dynamic-only threshold reduces accuracy below the baseline (77.19% vs 78.01%) on CIFAR-100, as early erasing removes useful information.
- The TET baseline drops 1.6 points below the baseline on DVS-CIFAR10, which the authors attribute to overfitting.
Limitations
- Hyperparameters (fixed threshold tau_f, kappa) were tuned on CIFAR-100 and show sensitivity.
- Evaluated only on image classification with ResNet backbones; no other architectures or tasks reported.
Experimental setup
Datasets: CIFAR-100 ImageNet DVS-CIFAR10
Benchmarks: CIFAR-100 ImageNet DVS-CIFAR10
Metrics: top-1 accuracy
Baselines: TET SLTT TTS GAC-SNN SlipReLU QCFS PSN MPBN LSG LM-H
Builds on: TET Adversarial erasing Self-erasing network LIF neuron
Closest papers by meaning
- Self-cross Feature based Spiking Neural Networks for Efficient Few-shot Learning (ICML 2025 · similarity 0.88)
- STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks (CVPR 2025 · similarity 0.88)
- Synergy Between the Strong and the Weak: Spiking Neural Networks are Inherently Self-Distillers (NeurIPS 2025 · similarity 0.87)
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.87)
- Rethinking Spiking Neural Networks from an Ensemble Learning Perspective (ICLR 2025 · similarity 0.87)
- Efficient Logit-based Knowledge Distillation of Deep Spiking Neural Networks for Full-Range Timestep Deployment (ICML 2025 · similarity 0.86)
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.