Abstract
Compared with artificial neural networks (ANNs), spiking neural networks (SNNs) are promising to explore the brain-like behaviors since the spikes could encode more spatio-temporal information. Although pre-training from ANN or direct training based on backpropagation (BP) makes the supervised training of SNNs possible, these methods only exploit the networks' spatial domain information which leads to the performance bottleneck and requires many complicated training skills. Another fundamental issue is that the spike activity is naturally non-differentiable which causes great difficulties in training SNNs. To this end, we build an iterative LIF model that is more friendly for gradient descent training. By simultaneously considering the layer-by-layer spatial domain (SD) and the timing-dependent temporal domain (TD) in the training phase, as well as an approximated derivative for the spike activity, we propose a spatio-temporal backpropagation (STBP) training framework without using any complicated technology. We achieve the best performance of multi-layered perceptron (MLP) compared with existing state-of-the-art algorithms over the static MNIST and the dynamic N-MNIST dataset as well as a custom object detection dataset. This work provides a new perspective to explore the high-performance SNNs for future brain-like computing paradigm with rich spatio-temporal dynamics.
Introduction
Spiking Neural Networks (SNNs) represent the "third generation" of neural networks, designed to emulate the brain's energy efficiency and temporal processing capabilities. Unlike traditional Artificial Neural Networks (ANNs) that communicate via continuous values, SNNs use discrete "spikes." Despite their potential, SNNs have historically lagged behind ANNs in performance due to two primary challenges: the "dead neuron" gradient problem caused by the non-differentiable nature of spikes, and the tendency of training methods to ignore the rich temporal dynamics inherent in spiking data.
This paper introduces Spatio-Temporal Backpropagation (STBP), a supervised learning framework that addresses these issues by treating the SNN as an iterative process unfolding across both space (layers) and time.
The STBP Framework
The core of the proposed method is the reformulation of the classic Leaky Integrate-and-Fire (LIF) neuron into a discrete-time iterative model. This transformation allows the neuron's state updates to be viewed similarly to a Recurrent Neural Network (RNN), facilitating differentiation.
The STBP algorithm propagates gradients through two distinct pathways:
- Spatial Domain (SD): The standard layer-to-layer flow found in traditional deep learning.
- Temporal Domain (TD): A time-step to time-step flow that captures the history of neuronal states.
By integrating these domains, the network can learn complex patterns that exist across time, rather than just static snapshots. Experiments showed that "Spatial-only" backpropagation (SDBP) performed significantly worse, highlighting that temporal information is the "missing link" for high-performance SNNs.
Overcoming Non-Differentiability
To solve the non-differentiable spike function problem—where the derivative is zero almost everywhere—the authors employ surrogate gradients. During the backward pass, the discontinuous step function is replaced with smooth, differentiable approximations.
The study evaluated four different curves: rectangular, polynomial, sigmoid, and Gaussian. A key finding was that the specific mathematical shape of the surrogate gradient is less critical than its ability to capture the nonlinear nature of the spike. This robustness simplifies the training process, removing the need for complex normalization or inhibition "tricks" used in previous SNN literature.
Experimental Results
The STBP framework was validated across three benchmarks, achieving state-of-the-art (SOTA) results for SNNs:
- Static MNIST: Achieved 98.89% accuracy, outperforming existing SNN-BP methods and ANN-to-SNN conversions.
- N-MNIST (Neuromorphic): On this dynamic dataset captured by event-based sensors, STBP reached 98.78% accuracy, demonstrating superior temporal processing.
- Custom Object Detection: The SNN (98.34%) slightly outperformed a standard non-spiking MLP (98.31%), proving that SNNs can match the performance of traditional ANNs when trained with spatio-temporal awareness.
Conclusion
STBP provides a streamlined, high-performance framework for training SNNs using standard gradient descent. While current evaluations focused on MLPs and small CNNs, future work will address scalability to larger architectures like ResNet and implementation on dedicated neuromorphic hardware to fully realize the energy-efficiency benefits of spiking dynamics.
Summary updated