Research map / Architectures and efficiency / Spiking networks
SPACE: SPike-Aware Consistency Enhancement for Test-Time Adaptation in 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 SPACE, a source-free, single-instance test-time adaptation method for spiking neural networks that maximizes consistency of spike-count-based local feature maps across AugMix views of one test sample.
Problem
SNNs are highly sensitive to distribution shifts, but existing test-time adaptation methods (designed for ANNs) rely on source data, batch statistics, or BN layers, and do not account for SNN spiking dynamics in the single-sample setting.
Approach
Adapt the SNN extractor with a single SGD step to maximize channel-wise similarity (softmax-normalized spatial distributions of time-aggregated spike counts) across augmented views of one test input, then predict with the adapted model; this uses spike-level feature consistency instead of output entropy or BN statistics.
Key results
- On CIFAR-10-C with SNN-VGG9 (level 5), SPACE reaches 71.03% average accuracy vs 66.57% No Adapt and 69.20% MEMO.
- On DVS Gesture-C with SNN-ConvLSTM, SPACE averages 65.91% vs 59.85% No Adapt and 60.23% MEMO.
- SPACE achieves the lowest GFLOPs (158.11) among compared TTA methods on CIFAR-10-C VGG9, versus 161.48 for MEMO and 318.07 for DeYO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply SPACE test-time adaptation | SNN-VGG9 · CIFAR-10-C, average | accuracy: 66.57 -> 71.03 | Helped |
| apply SPACE test-time adaptation | SNN-ResNet11 · CIFAR-10-C, average | accuracy: 65.96 -> 68.15 | Helped |
| apply SPACE test-time adaptation | SNN-VGG11 · CIFAR-100-C, average | accuracy: 40.63 -> 41.58 | Helped |
| apply SPACE test-time adaptation | SNN-VGG11 · Tiny-ImageNet-C, average | accuracy: 14.47 -> 16.48 | Helped |
| apply SPACE test-time adaptation | SNN-ConvLSTM · DVS Gesture-C, average | accuracy: 59.85 -> 65.91 | Helped |
| use 32 augmentations instead of 2 (ablation) | SPACE on SNN-VGG9 · CIFAR-10-C | accuracy: 54.60 -> 71.03 | Helped |
| increase AugMix strength from 1 to 10 (ablation) | SPACE on SNN-VGG9 · CIFAR-10-C, Gaussian Noise | accuracy: 77.98 -> 77.02 | Hurt |
| exceed AugMix semantic bounds (ablation) | SPACE on SNN-VGG9 · CIFAR-10-C, Gaussian Noise | accuracy: 77.98 -> 74.52 | Hurt |
| add kernel embedding (MMD) loss (ablation) | SPACE on SNN-VGG9 · CIFAR-10-C, average | accuracy: 71.03 -> 70.94 | Hurt |
| align average membrane potential instead of spike counts (ablation) | SPACE on SNN-VGG9 · CIFAR-10-C, average | accuracy: 71.03 -> 67.79 | Hurt |
| align spikes at each time step instead of total spike counts (ablation) | SPACE on SNN-VGG9 · CIFAR-10-C, average | accuracy: 71.03 -> 66.11 | Hurt |
| replace global feature-map similarity with local feature-map similarity (ablation) | SPACE · CIFAR-10-C, CIFAR-100-C, and Tiny-ImageNet-C | accuracy | Helped |
Surprising findings
- SITA performs worse than No Adapt on SNN-ResNet11 and CIFAR-10-C, suggesting BN-parameter adjustment disrupts spike information in SNNs.
- On CIFAR-10-C JPEG and some CIFAR-100-C corruptions, MEMO slightly outperforms SPACE.
Limitations
- Relies on backpropagation, which may limit use on resource-constrained or neuromorphic hardware.
- Performance drops when augmentation strength exceeds AugMix's semantic bounds.
- Some corruption types show MEMO slightly ahead of SPACE.
Experimental setup
Datasets: CIFAR-10-C CIFAR-100-C Tiny-ImageNet-C ImageNet-V2 ImageNet-R ImageNet-A DVS Gesture-C
Benchmarks: CIFAR-10-C CIFAR-100-C Tiny-ImageNet-C ImageNet-V2 ImageNet-R ImageNet-A DVS Gesture-C
Metrics: top-1 accuracy GFLOPs evaluation time Wilcoxon signed-rank p-value
Baselines: No Adapt MEMO SITA RoTTA DeYO
Builds on: MEMO SITA BNTT AugMix
Closest papers by meaning
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.89)
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.89)
- STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks (CVPR 2025 · similarity 0.88)
- Training High Performance Spiking Neural Network by Temporal Model Calibration (ICML 2025 · similarity 0.88)
- P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS (ICLR 2025 · similarity 0.88)
- SpikingSSMs: Learning Long Sequences with Sparse and Parallel Spiking State Space Models (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.