Disentangled Representation Learning Using ($\beta$-)VAE and GAN

Truly understanding a data might require identifying its generative factors. A concept that is more formally known as disentanglement. Classical approaches such as Principal Component Analysis (PCA) (Zietlow et al., 2021) has been developed for this purpose using linear algebra. In addition, VAE (Kingma and Welling, 2013) is a learning-based architecture that aims to represent the data in its disentangled latent space. In other words, VAEs were developed for learning a latent manifold that its axes align with independent generative factors of the data. Zietlow et al. (2021) argued that VAEs recover the nonlinear principal components of the data. In addition β-VAEs (Higgins et al., 2016) are a modified version of VAEs that when β > 1, weigh in more for disentanglement by sacrificing reconstruction quality. In this project, the goal is to explore the capacities of β-VAEs for learning a disentanglement representation, specifically to what degree the position of a moving object in the input frames can be encoded in the latent space.