Research map / Architectures and efficiency / Spiking networks
CLIF: Complementary Leaky Integrate-and-Fire Neuron 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 the Complementary LIF (CLIF) neuron, a parameter-free LIF variant with a complementary membrane potential that adds temporal gradient paths while keeping binary spikes, improving SG-trained SNN accuracy.
Problem
Surrogate-gradient-trained SNNs with LIF neurons lag ANN accuracy because temporal gradients vanish across timesteps, limiting benefit from longer BPTT horizons.
Approach
Adds a complementary membrane potential state m[t] that decays with the membrane potential, increases on each spike, and feeds a sigmoid term into the reset, creating extra temporal gradient paths without learnable parameters or non-binary outputs.
Key results
- CLIF with ResNet-18 reaches 96.69% on CIFAR-10 (8 timesteps) and 80.89% on CIFAR-100 with augmentation, slightly above the ReLU ANN (96.65% / 80.89%).
- On Tiny-ImageNet with VGG-13, CLIF reaches 64.13% at 6 timesteps versus 56.81% for ASGL with LIF.
- On DVS-Gesture, CLIF raises Spike-Driven-Transformer accuracy from 98.26% to 99.31% by swapping LIF for CLIF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace LIF with CLIF | Spiking-VGG11 · DVS-Gesture | accuracy: 97.57 -> 97.92 | Helped |
| Replace LIF with CLIF | Spike-Driven-Transformer · DVS-Gesture | accuracy: 98.26 -> 99.31 | Helped |
| Replace LIF with CLIF | Spiking-VGG11 · CIFAR10-DVS | accuracy: 78.05 -> 79.00 | Helped |
| Replace LIF with CLIF | VGGSNN · CIFAR10-DVS | accuracy: 84.90 -> 86.10 | Helped |
| Replace LIF with CLIF | ResNet-18 · CIFAR-10, augmented | accuracy: 96.45 -> 96.69 | Helped |
| Replace LIF with CLIF | ResNet-18 · CIFAR-100, augmented | accuracy: 80.58 -> 80.89 | Mixed |
| Increase timestep from 1 to 2 while using CLIF instead of LIF (ablation) | ResNet-18 · CIFAR-10 | accuracy: 92.7% | Helped |
| Extend BPTT gradient horizon from k=3 to k=6 (ablation) | LIF SNN · CIFAR-10 | accuracy: 94.24 -> 94.80 | Helped |
| Extend BPTT gradient horizon from k=4 to k=20 (ablation) | LIF SNN · DVS-Gesture | accuracy: 97.57 -> 97.57 | No clear effect |
Surprising findings
- Vanilla LIF accuracy peaks at T=32 and declines with more timesteps, while CLIF keeps improving.
- Only 2 timesteps of BPTT gradient are needed for LIF to reach most of its accuracy, suggesting longer temporal gradients contribute little.
Limitations
- Mathematical complexity of CLIF equations makes temporal gradient analysis hard, as the authors note.
- Gains on static datasets are modest (up to about 2%) and CLIF only slightly exceeds ANN accuracy in some settings.
- Analysis relies on a small 5-layer SNN and a few backbones, with limited hardware/compute details.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny-ImageNet DVS-Gesture CIFAR10-DVS
Benchmarks: CIFAR-10 CIFAR-100 Tiny-ImageNet DVS-Gesture CIFAR10-DVS
Metrics: classification accuracy firing rate
Baselines: LIF PLIF KLIF GLIF Dspike DSR PSN SML ASGL Online LTL Joint A-SNN ReLU ANN
Builds on: LIF neuron BPTT with surrogate gradient PLIF GLIF KLIF
Closest papers by meaning
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · similarity 0.90)
- TC-LIF: A Two-Compartment Spiking Neuron Model for Long-Term Sequential Modelling (AAAI 2024 · similarity 0.89)
- TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting (ICLR 2025 · similarity 0.88)
- The Expressive Leaky Memory Neuron: an Efficient and Expressive Phenomenological Neuron Model Can Solve Long-Horizon Tasks. (ICLR 2024 · similarity 0.87)
- Take A Shortcut Back: Mitigating the Gradient Vanishing for Training Spiking Neural Networks (NeurIPS 2024 · similarity 0.87)
- P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS (ICLR 2025 · similarity 0.87)
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.