One-Hot Multi-Level Leaky Integrate-and-Fire Spiking Neural Networks for Enhanced Accuracy-Latency Tradeoff

Pierre Abillama, Changwoo Lee, Andrea Bejarano-Carbo, Qirui Zhang, Dennis Sylvester, David T. Blaauw, Hun-Seok KimIEEE Access 2025 · DOI

TL;DR

This paper introduces a novel "one-hot multi-level leaky integrate-and-fire" (M-LIF) neuron model for Spiking Neural Networks (SNNs). This innovation allows SNNs to achieve a better balance between accuracy and energy efficiency, particularly in low-latency scenarios, by introducing an additional dimension of "spike lanes" to represent neuron outputs.

What Problem Does This Paper Address?

Spiking Neural Networks (SNNs) are promising for energy-efficient AI due to their event-driven nature, which can replace computationally expensive multiplications with additions. However, SNNs traditionally require multiple timesteps for inference, leading to increased latency and energy consumption. While methods exist to reduce latency to a single timestep, this often comes at the cost of significant accuracy degradation. The core challenge is finding an optimal trade-off between accuracy, latency, and energy efficiency. Existing approaches primarily focus on adjusting the number of timesteps, which has limitations.

Methods

The paper proposes a One-Hot Multi-Level Leaky Integrate-and-Fire (M-LIF) neuron model. The key innovation is the introduction of an "S" dimension, representing hidden layer outputs as a set of S binary-weighted "spike lanes." Crucially, the "one-hot" constraint ensures that only one of these spike lanes fires per timestep. This allows for a richer representation of neuron outputs beyond simple binary spikes, enabling finer-grained control over the accuracy-latency-energy trade-off without sacrificing the core energy-efficient nature of SNNs (i.e., no multiplications). The model is designed to be easily integrated into existing SNN training frameworks, particularly those using back-propagation through time (BPTT) with surrogate gradients.

Datasets

The authors evaluate their proposed M-LIF SNNs on several standard image classification datasets:

  • CIFAR10
  • CIFAR100
  • ImageNet
  • DVS-CIFAR10 (a dynamic vision dataset)

Key Insights and Ideas

  • Beyond Timesteps: The paper demonstrates that simply reducing timesteps isn't the only way to improve SNN efficiency. Introducing "spike lanes" (the S dimension) offers a new avenue for optimizing the accuracy-latency-energy trade-off.
  • One-Hot Constraint is Key: The "one-hot" nature of the spike lanes is critical. It ensures that even with multiple lanes, only one fires per timestep, maintaining computational simplicity and avoiding the multiplication overhead that other multi-level neuron approaches might introduce.
  • Bridging the Gap: The M-LIF model shows promise in closing the accuracy gap between SNNs and ANNs, especially in low-latency regimes where traditional SNNs struggle.
  • Energy Efficiency Gains: The proposed model achieves significant energy savings, particularly by reducing memory access energy due to fewer timesteps and potentially smaller activation bit-widths.

Key Results

  • Static Image Classification:
    • On ImageNet with VGG16, the one-hot M-LIF SNN achieved 71.05% accuracy with 3 timesteps and 3 spike lanes, outperforming conventional LIF SNNs by 2% and achieving 20x lower energy than a VGG16 ANN.
  • For spike-driven transformers, M-LIF SNNs boosted accuracy by up to 3% on ImageNet compared to LIF counterparts, with 4x less memory access energy due to fewer timesteps.
  • Dynamic Vision Tasks:
    • On DVS-CIFAR10, the M-LIF SNNs achieved 82.5% accuracy with only 3 timesteps and 4 spike lanes, a significant improvement over prior work that required more timesteps for comparable accuracy. This also resulted in a 3.45x reduction in computational energy.

Limitations and Future Work

  • Hardware Adaptability: While the model is designed to be adaptable to various hardware architectures, future work could focus on complete implementation and validation on ASICs or FPGAs to fully demonstrate its hardware advantages.
  • Broader Applications: The current experiments are primarily on image classification. The authors suggest expanding experiments to speech and text tasks to assess generalizability.
  • Robustness: Investigating robustness to noise and data distribution shifts is another important area for future research.
  • Training Techniques: Integrating techniques for training robust SNNs into the M-LIF framework could further enhance its capabilities.

Summary updated

BibTeX
@article{abillama2025one,
  author = {Pierre Abillama and Changwoo Lee and Andrea Bejarano-Carbo and Qirui Zhang and Dennis Sylvester and David T. Blaauw and Hun-Seok Kim},
  title = {{One-Hot Multi-Level Leaky Integrate-and-Fire Spiking Neural Networks for Enhanced Accuracy-Latency Tradeoff}},
  journal = {IEEE Access},
  year = {2025},
  doi = {10.1109/access.2025.3546508},
}
Figure