One-Class Classification

Abstract

This M.Sc. thesis report investigates the application of one-class classification techniques to complex high-dimensional data. The aim of a one-class classifier is to separate target data from non-target data, but only a dataset containing target data is available for training. The issue with high-dimensional data is that it is difficult to perform density estimation due to the `curse of dimensionality'. Most conventional method for one-class classification rely on density estimation.

This thesis focusses on the use of autoencoders and generative adversarial networks (GANs) for one-class classification problems involving image data. Autoencoders can learn encoding and decoding functions for samples from the target dataset. These encoding and decoding functions are, however, expected to not perform well for non-target samples, as they have never been seen during the training phase. This makes it possible to separate target and non-target data. For GANs, the discriminator is used to distinguish between target and non-target data.

Autoencoders and GANs are evaluated extensively in this report. Their behavior, desired parameters and strengths and weaknesses are evaluated by performing experiments. The main findings are that GANs do not perform well for one-class classification tasks, because of mode collapse and insufficient sampling of the non-target data. Even for extremely simple datasets these issues were observed. Autoencoders are shown to perform much better and behave according to the theoretical expectations.