Fully Spiking Variational Autoencoder

Spiking neural networks (SNNs) can be run on neuromorphic devices with ultra-high speed and ultra-low energy consumption because of their binary event-driven nature. Therefore, SNNs are expected to have various applications, including as generative models being running edge create high-quality images. In this study, we build a variational autoencoder (VAE) SNN enable image generation. VAE is known for its stability among models; recently, quality advanced. vanilla VAE, the latent space represented normal distribution, floating-point calculations required in sampling. However, not possible all features must time series data. constructed an autoregressive model, randomly selected samples from output sample variables. This allows variables follow Bernoulli process learning. Thus, Fully Variational Autoencoder where modules SNN. To best our knowledge, first only layers. We experimented several datasets, confirmed that it generate images same or better compared conventional ANNs. The code available at https://github.com/kamata1729/FullySpikingVAE.