Research map / Architectures and efficiency / Spiking networks
Ternary Spike: Learning Ternary Spikes 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 ternary {-1,0,1} spike neurons (and a trainable layer-wise amplitude variant re-parameterized back to standard ternary spikes at inference) to increase SNN information capacity while keeping multiplication-free computation.
Problem
Binary spike activations in spiking neural networks carry limited information, causing information loss and accuracy degradation compared with ANNs, and membrane potential distributions differ across layers yet are quantized to the same spike values.
Approach
Replace binary {0,1} spikes with ternary {-1,0,1} spikes (raising representable information from C×H×W bits to log2 3^(C×H×W)) and learn a layer-wise amplitude factor alpha, which is folded into weights via re-parameterization at inference so the network again uses normalized ternary spikes.
Key results
- Trainable ternary spike ResNet34 reaches 70.74% top-1 on ImageNet with 4 timesteps, versus 63.82% for binary spike baseline in the ablation and about 67.7% for prior SoTA.
- Ternary spike ResNet19 reaches 95.60% on CIFAR-10 with 2 timesteps, above prior best 95.51% (RMP-Loss).
- Trainable ternary spike ResNet19 reaches 79.80% on CIFAR10-DVS with 10 timesteps, versus 72.85% for Real Spike.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace binary spikes with ternary spikes (ablation) | ResNet18 and ResNet34 SNNs · ImageNet, 4 timesteps | accuracy: 61.07\% -> 66.90\%; 63.82\% -> 70.12\% | Helped |
| learn the layer-wise spike amplitude (ablation) | Ternary Spike SNN · ImageNet, 4 timesteps | accuracy: 66.90\% -> 67.68\%; 70.12\% -> 70.74\% | Helped |
| replace binary spikes with ternary spikes (ablation) | ResNet18 and ResNet34 SNNs · ImageNet, 2 timesteps | accuracy: 58.30\% -> 65.87\%; 62.81\% -> 69.48\% | Helped |
| learn the layer-wise spike amplitude (ablation) | Ternary Spike SNN · ImageNet, 2 timesteps | accuracy: 65.87\% -> 66.40\%; 69.48\% -> 69.51\% | Helped |
| learn the layer-wise spike amplitude | Ternary Spike · CIFAR-10, ResNet19, 2 timesteps | accuracy: 95.60\%$\pm 0.09$ -> 95.80\%$\pm 0.10$ | Helped |
| learn the layer-wise spike amplitude | Ternary Spike · CIFAR-100, ResNet19, 2 timesteps | accuracy: 79.66\%$\pm 0.08$ -> 80.20\%$\pm 0.10$ | Helped |
| learn the layer-wise spike amplitude | Ternary Spike · CIFAR10-DVS, ResNet19, 10 timesteps | accuracy: 78.40\%$\pm 0.21$ -> 79.80\%$\pm 0.16$ | Helped |
| use ternary spikes instead of binary spikes | ResNet18 and ResNet34 SNNs · ImageNet, 4 timesteps | energy: 50.14uJ -> 51.20uJ | Hurt |
Surprising findings
- Energy estimation shows only about 2.11% extra energy cost for ternary spikes over binary spikes (51.20uJ vs 50.14uJ).
Limitations
- Information-capacity argument is entropy-based and assumes uniform sample probabilities, which is a loose theoretical justification.
- Trainable variant trains with real-valued amplitude factor, so training cost and non-binary intermediate computation differ from standard SNN training.
- Energy estimate relies on assumed per-operation energy constants and counts sign operations separately.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet CIFAR10-DVS
Benchmarks: CIFAR-10 CIFAR-100 ImageNet CIFAR10-DVS
Metrics: top-1 accuracy energy consumption (uJ)
Baselines: TET RecDis-SNN RMP-Loss Dspike STBP-tdBN Real Spike DSR IM-Loss InfLoR-SNN SEW ResNet
Builds on: LIF neuron model TET RecDis-SNN InfLoR-SNN STBP-tdBN
Closest papers by meaning
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · similarity 0.89)
- Rethinking the Membrane Dynamics and Optimization Objectives of Spiking Neural Networks (NeurIPS 2024 · similarity 0.87)
- Take A Shortcut Back: Mitigating the Gradient Vanishing for Training Spiking Neural Networks (NeurIPS 2024 · similarity 0.87)
- QP-SNN: Quantized and Pruned Spiking Neural Networks (ICLR 2025 · similarity 0.87)
- ReverB-SNN: Reversing Bit of the Weight and Activation for Spiking Neural Networks (ICML 2025 · similarity 0.86)
- S$^2$NN: Sub-bit Spiking Neural Networks (NeurIPS 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.