Research map / Architectures and efficiency / Spiking networks
LM-HT SNN: Enhancing the Performance of SNN to ANN Counterpart through Learnable Multi-hierarchical Threshold Model
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 LM-HT, a learnable equidistant multi-threshold spiking neuron with a temporal-global information matrix that can be reparameterized into a single-threshold LIF model and initialized from QCFS-quantized ANNs for hybrid ANN-SNN training.
Problem
SNNs still lag behind ANNs in accuracy, and ANN-SNN conversion suffers severe degradation at low time-step latency; multi-threshold models lack rigorous theoretical grounding relative to vanilla SNNs and quantized ANNs.
Approach
An equidistant multi-threshold neuron (L levels over T steps) whose input current is mixed across time via a learnable T×T temporal-global information matrix and whose leak is learnable; it is mathematically equivalent to a vanilla IF/LIF neuron over LT steps and to LT-level quantized ANN activations under uniform input, enabling lossless reparameterization and QCFS-initialized hybrid training.
Key results
- Achieves 96.25% (ResNet-18) and 96.89% (ResNet-19) on CIFAR-10 and 79.33%/81.76% on CIFAR-100 with only 2 time-steps.
- Reaches 70.90% top-1 on ImageNet-1k with ResNet-34 at 2 time-steps, 3.38% above GLIF at 4 steps.
- In hybrid training on CIFAR-100 with ResNet-20 (QCFS T_q=8), LM-HT (L=2) at 2 steps reaches 67.08% vs 19.76% for the QCFS-converted SNN (+32.91%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add learnable multi-hierarchical thresholds (L=2) | SNN direct training with ResNet-19 · CIFAR-100 | accuracy: 78.28 -> 81.76 | Helped |
| replace vanilla single-threshold setting (L=1, T=4) with L=2, T=2 (ablation) | LM-HT, ResNet-18 · ImageNet-1k subset | accuracy: 76.22 -> 80.52 | Helped |
| add T-GIM (ablation) | LM-HT, ResNet-18 · ImageNet-1k subset | accuracy: 80.52 -> 80.56 | Helped |
| add T-GIM (ablation) | LM-HT, ResNet-34 · ImageNet-1k subset | accuracy: 82.18 -> 82.72 | Helped |
| add T-GIM (ablation) | LM-HT, ResNet-18 · ImageNet-1k subset | SOPs: 1.08 -> 0.73 | Helped |
| add T-GIM (ablation) | LM-HT, ResNet-34 · ImageNet-1k subset | SOPs: 2.42 -> 1.72 | Helped |
| reparameterize the multi-threshold model as a single-threshold model (ablation) | LM-HT, VGG-13 · ImageNet-1k | accuracy: 61.64 -> 61.66 | No clear effect |
| reparameterize the multi-threshold model as a single-threshold model (ablation) | LM-HT, ResNet-18 · ImageNet-1k | accuracy: 64.44 -> 64.50 | No clear effect |
| replace QCFS conversion with LM-HT hybrid training | VGG-16 SNN · CIFAR-100, Tq=4, 2 steps | accuracy: 63.33 -> 75.97 (+6.27) | Helped |
| replace QCFS conversion with LM-HT hybrid training | ResNet-20 SNN · CIFAR-100, Tq=8, 2 steps | accuracy: 19.76 -> 67.08 (+32.91) | Helped |
Limitations
- Paper's own checklist states no limitations discussed.
- No error bars or significance tests; experiments use a shared random seed.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-200 ImageNet-1k CIFAR10-DVS
Benchmarks: CIFAR-10 CIFAR-100 ImageNet-200 ImageNet-1k CIFAR10-DVS
Metrics: top-1 accuracy synaptic operations (SOPs) energy (mJ)
Baselines: STBP-tdBN Dspike TET SLTT GLIF RMP-Loss MBPN SEW ResNet ASGL QCFS DCT Online-LTL
Builds on: STBP QCFS multi-threshold neuron model tdBN
Closest papers by meaning
- Enhancing Training of Spiking Neural Network with Stochastic Latency (AAAI 2024 · similarity 0.90)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.90)
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · similarity 0.89)
- Efficient ANN-Guided Distillation: Aligning Rate-based Features of Spiking Neural Networks through Hybrid Block-wise Replacement (CVPR 2025 · similarity 0.89)
- Enhanced Self-Distillation Framework for Efficient Spiking Neural Network Training (NeurIPS 2025 · similarity 0.89)
- Efficient ANN-SNN Conversion with Error Compensation Learning (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.