Spatial Variational Auto-Encoding via Matrix-Variate Normal Distributions

TL;DR

This paper introduces spatial variational auto-encoders that utilize matrix-variate normal distributions for explicit spatial encoding in latent feature maps, enhancing structural and spatial information capture.

Contribution

It proposes a novel spatial VAE framework using matrix-variate normal distributions, including a low-rank variant to improve spatial dependency modeling and parameter efficiency.

Findings

Spatial VAEs outperform traditional VAEs in capturing spatial information.

Low-rank MVN distributions reduce parameters while maintaining performance.

Experimental results demonstrate improved structural representation.

Abstract

The key idea of variational auto-encoders (VAEs) resembles that of traditional auto-encoder models in which spatial information is supposed to be explicitly encoded in the latent space. However, the latent variables in VAEs are vectors, which can be interpreted as multiple feature maps of size 1x1. Such representations can only convey spatial information implicitly when coupled with powerful decoders. In this work, we propose spatial VAEs that use feature maps of larger size as latent variables to explicitly capture spatial information. This is achieved by allowing the latent variables to be sampled from matrix-variate normal (MVN) distributions whose parameters are computed from the encoder network. To increase dependencies among locations on latent feature maps and reduce the number of parameters, we further propose spatial VAEs via low-rank MVN distributions. Experimental results show that the proposed spatial VAEs outperform original VAEs in capturing rich structural and spatial information.