SpikeGAN: Spike-Based Generative Adversarial Networks

Spiking neural networks (SNNs) offer a route to energy-efficient generative modeling through event-driven computation, but adversarial learning must account for discrete spike representations and temporal output variability. We introduce SpikeGAN, which jointly trains an auxiliary artificial neural network (ANN) generator and an SNN generator with a shared critic. Continuous ANN samples shape the critic's representation and provide distribution-level guidance to the SNN. A temporal objective supplies direct adversarial feedback to each SNN output step, so intermediate generated samples receive their own training signal. We derive a sufficient covariance condition under which per-step supervision has lower gradient variance than average-output supervision. Experiments on CIFAR-10, CelebA, and LSUN bedroom evaluate generation quality and efficiency, while controlled ablations and multi-seed diagnostics assess the training components. On CIFAR-10, SpikeGAN achieves an FID of 15.19 with four time steps and 4.30M generator parameters, compared with the reported SDDPM result of 16.89 with 63.61M parameters, a parameter-count ratio of approximately $14.8\times$. These results support joint distribution guidance and temporal supervision for compact spike-based generative modeling.