Research map / Architectures and efficiency / Spiking networks
Are Conventional SNNs Really Efficient? A Perspective from Network Quantization
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.
Shows that T-step SNNs are equivalent to T-bit quantized ANNs and proposes the "Bit Budget" metric (T·bw·bs) with S-ACE/NS-ACE to fairly estimate SNN energy and allocate bits between time steps, spike precision, and weights.
Problem
SNN energy efficiency is usually estimated with synaptic operations (SOPs) and compared against unoptimized ANNs, which skews comparisons; the relation between SNN time steps and quantized ANN bit-widths is unclear.
Approach
Unifies SNN time steps and quantized activation bit-widths as equivalent representations, defines the Bit Budget (T × weight bit-width × spike bit-width) as a per-synapse cost measure, and uses it to guide allocating bits between time steps and multi-bit spike patterns plus weight quantization.
Key results
- On ImageNet, a 1-bit-weight Spikformer-8-512 with 4-bit weights and 8-bit spikes (T=1) reaches 79.37% accuracy at 1.86-14.84M params range, outperforming the 16/1/4 Spikformer-8-512 (73.38%) with the same bit budget family.
- On CIFAR10/100, the 4/4/1 allocation achieves 96.84% / 80.13%, versus 95.51% / 78.21% for the 16/1/4 SNN baseline with much lower S-ACE (12.32G vs 59.10G).
- FPGA implementation on xck26 for CIFAR10 with 8/4/1 bit allocation achieves 343.29 fps at 96.41% accuracy, with software and hardware results congruent.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Allocate more bits to spike patterns and fewer to time steps for static-image SNNs (ablation) | Quantized SNNs · ImageNet, CIFAR10, CIFAR100 | accuracy: 96.50 vs 95.94; 80.71 vs 78.77 | Helped |
| Increase spike-pattern bits while reducing time steps (ablation) | Quantized SNNs · CIFAR100 | accuracy: 80.71 vs 78.77 | Helped |
| Increase time steps rather than spike-pattern bit-width (ablation) | Quantized SNNs · CIFAR10 | accuracy: 93.91 vs 95.00 | Hurt |
| Use 1 / 2 / 8 rather than 1 / 1 / 16 bit allocation (ablation) | Quantized SNN · DVS Gesture | accuracy (%): 96.67 (+0.00) -> 98.48 (+1.81) | Helped |
| Use 1 / 4 / 4 rather than 1 / 1 / 16 bit allocation (ablation) | Quantized SNN · N-Caltech101 | accuracy (%): 79.77 (+0.00) -> 82.64 (+2.87) | Helped |
| Use 1 / 4 / 4 rather than 1 / 1 / 16 bit allocation (ablation) | Quantized SNN · DVS-CIFAR10 | accuracy (%): 79.8 (+0.00) -> 63.1 (-16.7) | Hurt |
| Use 1 / 16 / 1 rather than 1 / 1 / 16 bit allocation (ablation) | Quantized SNN · DVS-CIFAR10 | accuracy (%): 79.8 (+0.00) -> 35.8 (-44.0) | Hurt |
| Quantize weights and spike patterns to 8 / 4 / 1 | Spikformer-8-512 · ImageNet | accuracy (%): 73.38 -> 79.37 | Helped |
| Quantize weights and spike patterns to 8 / 8 / 1 | SEW-ResNet-34 · ImageNet | accuracy (%): 67.04 -> 70.13 | Helped |
| Allocate more bits to spike patterns instead of time steps (ablation) | SNNs · CIFAR100 and N-Caltech101 | accuracy | Mixed |
Surprising findings
- Allocating bits to spike precision rather than time steps improves accuracy on static images, and 1-bit weight models still perform well.
- On neuromorphic data, adding spike bit-width too far (1/16/1) degrades accuracy, showing a balance is needed between spike bits and time steps.
Limitations
- Quantization technique itself is not contributed; relies on standard symmetric quantization.
- Energy estimates are analytical and validated only with a linear fit on FPGA, not on real neuromorphic chips.
- Paper states the goal is not to set new SNN SOTA, and results for neuromorphic datasets are limited to a single model.
Experimental setup
Datasets: ImageNet CIFAR10 CIFAR100 CIFAR10-DVS DVS128 Gesture N-Caltech101
Benchmarks: ImageNet CIFAR10 CIFAR100 CIFAR10-DVS DVS128 Gesture N-Caltech101
Metrics: top-1 accuracy parameters (M) SOPs (G) S-ACE (G) NS-ACE (G) fps firing rate
Baselines: Hybrid training TET Spiking ResNet tdBN SEW ResNet Spikformer
Builds on: Spikformer SEW ResNet surrogate gradient synaptic operations (SOPs)
Closest papers by meaning
- QP-SNN: Quantized and Pruned Spiking Neural Networks (ICLR 2025 · similarity 0.92)
- S$^2$NN: Sub-bit Spiking Neural Networks (NeurIPS 2025 · similarity 0.91)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.90)
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.89)
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.89)
- Rethinking the Membrane Dynamics and Optimization Objectives of Spiking Neural Networks (NeurIPS 2024 · similarity 0.89)
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.