Abstract
Spiking Neural Networks (SNNs) offer a biologically plausible framework forenergy-efficient neuromorphic computing. However, it is a challenge to trainSNNs due to their non-differentiability, efficiently. Existing gradientapproximation approaches frequently sacrifice accuracy and face deploymentlimitations on edge devices due to the substantial computational requirementsof backpropagation. To address these challenges, we propose a Forward-Forward(FF) based gradient approximation-free training framework for Spiking NeuralNetworks, which treats spiking activations as black-box modules, therebyeliminating the need for gradient approximation while significantly reducingcomputational complexity. Furthermore, we introduce a class-aware complexityadaptation mechanism that dynamically optimizes the loss function based oninter-class difficulty metrics, enabling efficient allocation of networkresources across different categories. Experimental results demonstrate thatour proposed training framework achieves test accuracies of 99.58%, 92.13%, and75.64% on the MNIST, Fashion-MNIST, and CIFAR-10 datasets, respectively,surpassing all existing FF-based SNN approaches. Additionally, our proposedmethod exhibits significant advantages in terms of memory access andcomputational power consumption.