Spiking Neural Networks (SNNs) are highly attractive for edge computing and neuromorphic hardware due to their extreme energy efficiency. Unlike traditional Artificial Neural Networks (ANNs) that communicate via continuous values, SNNs process information using discrete, binary events called spikes.
However, this biological realism introduces a fundamental mathematical dilemma. A spiking neuron accumulates input signals into a membrane potential; once this potential crosses a specific threshold, the neuron fires a spike. Mathematically, this thresholding operation is represented by a step function. The derivative of a step function is zero everywhere except at the exact threshold, where it is undefined.
Because standard deep learning relies on backpropagation—which uses the chain rule of calculus to propagate continuous error gradients backward through layers—this non-differentiable activation function causes gradients to vanish instantly. As noted by A. Gruening et al. in Spiking Neural Networks: Principles and Challenges, this mathematical discontinuity prevents the direct application of gradient-based optimization.
To resolve this bottleneck, researchers have developed three primary paradigms: surrogate gradients, ANN-to-SNN conversion, and gradient-free local learning rules.
Surrogate Gradients: Softening the Backward Pass
The most prominent strategy for training SNNs directly is the surrogate gradient method. The core intuition is to split the network's behavior into two distinct modes: a discrete forward pass and a continuous backward pass.
During the forward pass, the network retains its hard, non-differentiable threshold functions. This preserves the sparse, energy-efficient properties of the network during inference. During the backward pass, however, the non-differentiable step function is replaced by a smooth, differentiable approximation (such as a sigmoid, piece-wise linear, or Gaussian function) solely for the purpose of calculating gradients.
Forward Pass (Discrete Step): Backward Pass (Smooth Surrogate):
▲ ▲
│ ┌─── │ _.._
│ │ │ .' '.
───────┴───┴──────► ───────┴────────►
Threshold V_th Threshold V_th
Why does this work? Optimization algorithms like Stochastic Gradient Descent (SGD) do not require mathematically perfect gradients to find a local minimum; they merely require an error signal that points in a generally correct direction.
Early foundations for this temporal backpropagation were laid by Sander Bohte et al. in SpikeProp: backpropagation for networks of spiking neurons by linearizing the thresholding function around the firing time. Modern frameworks scale this concept across space and time. For instance, Yujie Wu et al. introduced Spatio-Temporal Backpropagation (STBP) in Spatio-Temporal Backpropagation for Training High-performance Spiking Neural Networks, which treats the SNN as an unrolled recurrent network and applies surrogate gradients across both network layers and time steps.
To make this process computationally viable for deep networks, Qingyan Meng et al. proposed Spatial Learning Through Time (SLTT) in Training much deeper spiking neural networks with a small number of time-steps, which discards negligible temporal gradients to slash memory overhead during surrogate backpropagation.
ANN-to-SNN Conversion: Sidestepping the Gradient
An alternative paradigm avoids direct training altogether by leveraging the optimization maturity of traditional deep learning. In ANN-to-SNN conversion, a standard network with continuous activation functions (typically ReLU) is trained using standard backpropagation. Afterward, its weights and biases are mapped onto an SNN architecture.
The underlying intuition relies on rate coding: mapping the continuous output magnitude of an ANN neuron to the firing frequency of a spiking neuron over a given time window. If an ANN neuron outputs a high value, the corresponding SNN neuron is calibrated to fire frequently.
ANN Continuous Activation (ReLU) SNN Rate Coding
▲ ▲
│ / │ │ │ │ │ │ │
│ / │ │ │ │ │ │ │
──────────┴─┴─────► ──────────────►
Input Time Window
However, simple mapping often introduces conversion errors due to the temporal quantization of spikes. To minimize this accuracy loss, the authors of A Free Lunch From ANN: Towards Efficient, Accurate Spiking Neural Networks Calibration developed layer-wise calibration techniques that adjust weights, biases, and initial membrane potentials to align the two domains precisely.
Similarly, Jianhao Ding et al. proposed an optimal conversion strategy in Optimal ANN-SNN Conversion for Fast and Accurate Inference in Deep Spiking Neural Networks by introducing a trainable Rate Norm Layer during the ANN phase. To ensure the ANN behaves more like an SNN prior to conversion, Zhanglu Yan et al. introduced Clamped and Quantized (CQ) Training in Near Lossless Transfer Learning for Spiking Neural Networks, forcing the source ANN to mimic the discrete, restricted nature of an SNN from the start.
Alternative Learning Rules: Moving Beyond Backpropagation
A third approach argues that forcing backpropagation onto SNNs is fundamentally mismatched with neuromorphic hardware constraints. Instead, these methods utilize gradient-free or biologically inspired local learning rules.
Backpropagation requires symmetric weight transport to pass error signals backward through the exact same channels used in the forward pass. To bypass this requirement, Dongcheng Zhao et al. developed GLSNN in GLSNN: A Multi-Layer Spiking Neural Network Based on Global Feedback Alignment and Local STDP Plasticity. This framework combines global feedback alignment—which uses fixed, random pathways to distribute error signals—with local Spike-Timing-Dependent Plasticity (STDP) to update weights based on the relative timing of pre- and post-synaptic spikes.
For on-chip deployment, Jeong-Jun Lee et al. introduced Spike-Train Level Direct Feedback Alignment (ST-DFA) in Spike-Train Level Direct Feedback Alignment: Sidestepping Backpropagation for On-Chip Training of Spiking Neural Nets. By replacing layer-by-layer error propagation with direct, random feedback signals sent straight from the output layer to hidden layers, the architecture sidesteps the non-differentiability barrier entirely, significantly reducing hardware resource overhead and training latency.