Use of Student's t-Distribution for the Latent Layer in a Coupled Variational Autoencoder

TL;DR

This paper introduces a coupled variational autoencoder that employs a Student's t-distribution in the latent layer and a coupled loss function, enhancing the accuracy and robustness of generated MNIST images, especially in handling outliers.

Contribution

It proposes a novel combination of Student's t-distribution for the latent space and a coupled logarithmic loss function to improve generative model performance.

Findings

Improved accuracy in MNIST image generation

Enhanced robustness to outliers

Better decisiveness in generated images

Abstract

A Coupled Variational Autoencoder, which incorporates both a generalized loss function and latent layer distribution, shows improvement in the accuracy and robustness of generated replicas of MNIST numerals. The latent layer uses a Student's t-distribution to incorporate heavy-tail decay. The loss function uses a coupled logarithm, which increases the penalty on images with outlier likelihood. The generalized mean of the generated image's likelihood is used to measure the performance of the algorithm's decisiveness, accuracy, and robustness.