An Analysis on Disentanglement in Machine Learning

Abstract

Learnt representations by Deep autoencoders is not capable of decomposing the complex information into simple notion. In other words, attributes of samples are entangled in the basis vectors spanning the learned space. This leads to significant errors in deep learning algorithms. In order to avoid these errors, it is necessary to separate the feature space according to the common features shared between classes and to define a simple subspace for each feature. This approach has led to the birth of a new paradigm in Machine Learning, called disentanglement.Roughly, disentangled models can be defined as models that can independently learn the different components of the probability density function that produces the dataset in the feature space. Unfortunately, it is not always possible to learn these models. For this reason, there is still no easily applicable mathematical definition of disentanglement in the literature. In this study, a mathematical definition of the concept of disentanglement will be made and methods and metrics related to this approach will be discussed.