Disentangled conditional variational autoencoder for unsupervised anomaly detection
Loading...
Date
2022-12-05
Authors
Neloy, Asif Ahmed
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The goal of efficient anomaly or outlier detection is to learn the hidden representation of the data by identifying independent factors and minimizing information loss. Variational Autoencoder (VAE) and its extensions have shown great promise to learn the data distribution. In this manuscript-based thesis, I propose a novel architecture of generative framework to investigate the following objectives: (a) Effectively learn disentangled representations of data by optimizing total correlation (TC) loss. (b) Minimize information loss using mutual information theory for better reconstruction ability and appropriate sample generation. (c) Address sample reconstruction error vs. reconstruction quality trade-offs. In the first manuscript, I review the architectures of autoencoders divided into three main categories: classical, variational, and regularized autoencoders. Then, I present their mathematical foundation and explore their ability to detect anomalies, reconstruct samples and learn latent factors. In the second manuscript, I propose Disentangled Conditional Variational Autoencoder dCVAE, which combines the frameworks of 𝛽-VAE, conditional variational autoencoder (CVAE), and the principle of total correlation (CorEx). Through experiments, I show that the accuracy of anomaly detection methods can be improved while learning disentangled factors and minimizing information loss. This can be done by connecting multivariate information theory and regularizing the posterior-variant of VAE. Finally, I conclude this thesis by discussing limitations, and I give a brief overview of future research directions for VAE architectures in high-dimensional image datasets.
Description
Keywords
anomaly detection, unsupervised machine learning, disentangled representations, mutual information theory