Research map / Architectures and efficiency / Spiking networks
ReverB-SNN: Reversing Bit of the Weight and Activation for 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 ReverB-SNN, which pairs real-valued spike activations with binary weights (plus a learnable per-channel scale folded back at inference) to improve directly trained SNN accuracy while keeping addition-only computation.
Problem
Binary spike activations in spiking neural networks carry limited information, reducing accuracy, while existing fixes (ternary, integer, real-valued spikes) sacrifice the multiplication-free property.
Approach
Reverse the usual precision assignment: use real-valued spike activations with binary {-1,1} weights so that multiplications remain additions; add a learnable channel-wise amplitude alpha during training and fold it into the previous layer's activations via re-parameterization at inference.
Key results
- ResNet34 with 4 timesteps reaches 70.91% top-1 on ImageNet with the learnable variant, reported as 3.22% above other SNN SoTA.
- ResNet19 reaches 96.62% on CIFAR-10 and 78.46% on CIFAR-100 at 2 timesteps with the learnable variant.
- ResNet19 reaches 80.50% on CIFAR10-DVS with the learnable variant, versus 80.00% for SSCL as the best prior result in the table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace binary spike activations with real-valued spikes and use binary weights (ablation) | ResNet20 SNN · CIFAR-10, 2 timesteps | top-1 accuracy: 92.80% -> 94.14% | Helped |
| Replace binary spike activations with real-valued spikes and use binary weights (ablation) | ResNet20 SNN · CIFAR-100, 2 timesteps | top-1 accuracy: 70.18% -> 72.54% | Helped |
| Add learnable binary-weight amplitude (ablation) | ReverB ResNet20 · CIFAR-10, 2 timesteps | top-1 accuracy: 94.14% -> 94.45% | Helped |
| Add learnable binary-weight amplitude (ablation) | ReverB ResNet20 · CIFAR-100, 2 timesteps | top-1 accuracy: 72.54% -> 72.95% | Helped |
| Add learnable binary-weight amplitude (ablation) | ReverB ResNet20 · CIFAR-10, 4 timesteps | top-1 accuracy: 94.55% -> 94.96% | Helped |
| Add learnable binary-weight amplitude (ablation) | ReverB ResNet20 · CIFAR-100, 4 timesteps | top-1 accuracy: 72.93% -> 73.28% | Helped |
| Use real-valued spikes with binary weights instead of standard SNN activations and weights | ResNet20 SNN · CIFAR-10, 2 timesteps | energy: 49.73uJ -> 49.99uJ | Hurt |
| Use real-valued spikes with binary weights | SNN · CIFAR-10, ResNet19, 2 timesteps | top-1 accuracy: 95.53% -> 96.39%$\pm 0.11$ | Helped |
| Use learnable binary-weight amplitude | ReverB SNN · ImageNet, ResNet34, 4 timesteps | top-1 accuracy: 70.74%$\pm 0.13$ -> 70.91%$\pm 0.13$ | Helped |
| Use learnable binary-weight amplitude | ReverB SNN · CIFAR10-DVS, ResNet19, 10 timesteps | top-1 accuracy: 80.30%$\pm 0.20$ -> 80.50%$\pm 0.10$ | Helped |
Surprising findings
- Energy cost of ReverB-SNN is reported as only 0.52% higher than the vanilla SNN despite real-valued spikes.
Limitations
- Real-valued spikes mean activations are no longer strictly binary, so the energy-efficiency argument relies on an analytical estimate rather than hardware measurement.
- Gains on CIFAR-10 for ReverB are modest relative to the vanilla baseline at 4 timesteps in the ablation (about 0.7 points).
- Energy estimate uses assumed per-operation costs and a sparsity-based SOP count.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet CIFAR10-DVS
Benchmarks: CIFAR-10 CIFAR-100 ImageNet CIFAR10-DVS
Metrics: top-1 accuracy FLOPs SOPs energy
Baselines: Real Spike Ternary spike TET RecDis-SNN Dspike GLIF DSR SEW ResNet SSCL Vanilla SNN
Builds on: STBP EGRU Straight-through estimator Leaky integrate-and-fire neuron
Closest papers by meaning
- Memory-Efficient Reversible Spiking Neural Networks (AAAI 2024 · similarity 0.91)
- Take A Shortcut Back: Mitigating the Gradient Vanishing for Training Spiking Neural Networks (NeurIPS 2024 · similarity 0.90)
- High-Performance Temporal Reversible Spiking Neural Networks with $\mathcal{O}(L)$ Training Memory and $\mathcal{O}(1)$ Inference Cost (ICML 2024 · similarity 0.89)
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · similarity 0.89)
- Are Conventional SNNs Really Efficient? A Perspective from Network Quantization (CVPR 2024 · similarity 0.88)
- Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation (ICML 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.