An Autoencoder-based Spiking Representation for Images

Autoencoders effectively extract low-dimensional features in artificial neural networks (ANNs) but remain scarcely explored for spiking neural networks (SNNs) which enable low-power hardware implementation. This paper presents a spiking autoencoder (SAE) trained by surrogate gradient backpropagation. We implement an encoder-decoder architecture and optimize it to reconstruct images from a compressed spiking representation. Experiments on MNIST and Fashion-MNIST show the SAE accurately reconstructs inputs and compresses them into binary spike codes. Our work demonstrates successful SAE training using surrogate gradients and shows potential for exploiting autoencoders for efficient neural coding with SNNs.