Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Self-cross Feature based Spiking Neural Networks for Efficient Few-shot 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 SSCF, a spiking few-shot learning framework combining a self-feature extractor and a cross-feature contrastive module on a VGGSNN backbone, trained with TET loss plus InfoNCE loss.
Problem
Few-shot learning with DNNs is computationally expensive and energy-intensive, and existing SNN few-shot methods struggle to capture complex spatiotemporal features and perform accurate cross-class comparison.
Approach
Combines a self-correlation feature extractor (SFE) on spiking backbone features with a cross-correlation attention module (CFC) that matches query and support features, optimized with a TET loss and an InfoNCE contrastive loss in a weighted sum.
Key results
- On N-Omniglot 5w5s, SSCF with VGGSNN reaches 98.9% (T=8), versus 90.7% for Siamese/MAML baselines reported in the table.
- On CUB 5-way 1-shot, SSCF reaches 76.27%, below RENet's 79.49% but above ProtoNet's and others' reported numbers.
- Estimated energy for SSCF at T=2 is 1.849 mJ versus 22.264 mJ for RENet (about 8.3% of the ANN energy).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SSCF with VGGSNN rather than SCNN | SSCF · N-Omniglot 20-way 1-shot | accuracy: 67.8±0.5 -> 84.4±0.6 | Helped |
| Use SSCF instead of Siamese | SCNN · N-Omniglot 20-way 1-shot | accuracy: 49.8±1.3 -> 67.8±0.5 | Helped |
| Increase time steps from T=4 to T=12 | SSCF with VGGSNN · N-Omniglot 5-way 1-shot | accuracy: 94.3±0.8 -> 95.3±0.6 | Helped |
| Increase time steps from T=4 to T=12 | SSCF with VGGSNN · N-Omniglot 5-way 5-shot | accuracy: 98.6±0.3 -> 98.8±0.2 | No clear effect |
| Use SSCF rather than RENet | CUB · CUB 5-way 1-shot | accuracy: 79.49±0.44 -> 76.27±0.46 | Hurt |
| Use SSCF rather than RENet | miniImageNet · miniImageNet 5-way 5-shot | accuracy: 82.58±0.30 -> 75.61±0.34 | Hurt |
| Add SFE (ablation) | SSCF · N-Omniglot 5-way 1-shot | accuracy: 93.17 -> 94.34 | Helped |
| Add SFE (ablation) | SSCF · CUB 5-way 1-shot | accuracy: 71.84 -> 76.27 | Helped |
| Add CFC (ablation) | SSCF · N-Omniglot | accuracy: 93.17 -> 94.34 | Helped |
| Add CFC (ablation) | SSCF · CUB 5-way 1-shot | accuracy: 72.20 -> 76.27 | Helped |
| Use InfoNCE loss instead of CE loss (ablation) | SSCF · CUB with noise rate 0.4 | accuracy: 40.649 -> 43.699 | Helped |
| Use InfoNCE loss instead of CE loss (ablation) | SSCF · CUB with noise rate 0.8 | accuracy: 32.725 -> 36.485 | Helped |
Surprising findings
- On miniImageNet 5-way 5-shot, SSCF (75.61%) is well below several ANN baselines, and the 5w1s result (60.97%) trails RENet (67.60%).
Limitations
- Accuracy on static datasets (CUB, miniImageNet) remains below the best ANN methods such as RENet.
- The energy comparison uses theoretical SOP/FLOP estimates rather than measured hardware energy.
- Several results tables are inconsistent (e.g., text vs table values, unexplained settings), which weakens confidence in the reported numbers.
Experimental setup
Datasets: N-Omniglot CUB-200-2011 miniImageNet
Benchmarks: N-Omniglot CUB-200-2011 miniImageNet
Metrics: accuracy SOPs FLOPs energy (mJ)
Baselines: Siamese network MAML ProtoNet RENet DeepEMD FEAT Knowledge-Transfer OWOML
Builds on: temporal efficient training InfoNCE leaky integrate-and-fire
Closest papers by meaning
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · similarity 0.89)
- Enhancing Representation of Spiking Neural Networks via Similarity-Sensitive Contrastive Learning (AAAI 2024 · similarity 0.88)
- FSTA-SNN:Frequency-Based Spatial-Temporal Attention Module for Spiking Neural Networks (AAAI 2025 · similarity 0.88)
- Enhanced Self-Distillation Framework for Efficient Spiking Neural Network Training (NeurIPS 2025 · similarity 0.88)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.88)
- Towards More Discriminative Feature Learning in SNNs with Temporal-Self-Erasing Supervision (AAAI 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.