Yuechao Wang, Yizhou Jiang, Yihan Li, Yuqian Liu, Tianren Zhang, Ying Fang, Feng Chen. Learning to be Lazy: Dynamic Timestep Learning in Spiking Neural NetworksJ. Machine Intelligence Research, 2026, 23(6): 1355-1366. DOI: 10.1007/s11633-025-1608-7
Citation: Yuechao Wang, Yizhou Jiang, Yihan Li, Yuqian Liu, Tianren Zhang, Ying Fang, Feng Chen. Learning to be Lazy: Dynamic Timestep Learning in Spiking Neural NetworksJ. Machine Intelligence Research, 2026, 23(6): 1355-1366. DOI: 10.1007/s11633-025-1608-7

Learning to be Lazy: Dynamic Timestep Learning in Spiking Neural Networks

  • Spiking neural networks (SNNs) have attracted significant interest due to their biological plausibility and energy efficiency. Conventional SNN implementations mostly employ rate coding strategies with fixed simulation timestep, referred to as latency, entailing increased inference latency and power consumption for higher performance. Thus, in this work, we propose to seek a minimum latency for different examples through training to address this tradeoff. Our proposed target termination strategy regards the timestep for different examples as a variable parameter and minimizes it for earlier predictions. Prior to simulation termination, the temporary accumulated output at each timestep is integrated into the loss function, ensuring the correctness of early results. Additionally, we introduce a temporal regularization term to ensure the monotonicity of the loss function over time, implicitly reducing latency through optimization. We provide a theoretical foundation for our method and demonstrate its efficacy through empirical evaluations on CIFAR and ImageNet, revealing the potential of dynamic timestep strategies in enhancing SNNs efficiency.
  • loading

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return