Abstract
Abstract
Disentangled representation learning (DRL) is the problem of obtaining a representation where the underlying sources of variation generating the data are independently
captured. A pioneering DRL method is Beta Variational Auto Encoder (β-VAE), which extends the infamous VAE with a hyperparameter β weighting the optimization of the disentanglement term formulated as a Kullback-Leibler Divergence term.
In this thesis, we make three contributions to DRL. First, we experimentally analyze β-VAE and show that it is highly sensitive to the choice of β for finding an optimal
balance between the order of disentanglement and the representation capacity measured through the quality of reconstruction.
Second, to address this hyperparameter sensitivity issue of β-VAE, we propose two novel methods: Learnable VAE (L-VAE), and Dimensionwise Learnable VAE (dL-VAE). L-VAE mitigates the limitations of β-VAE by learning the relative weights of the disentanglement and the reconstruction terms in the loss function to dynamically
control the trade-off between disentanglement and reconstruction. The weight of the loss terms and the parameters of the model architecture are learned concurrently, eliminating the complexity of hyperparameter optimization. dL-VAE, on the other hand, learns dimension-wise weighting hyperparameters, formulating the disentanglement-
reconstruction trade-off per dimension of the representation. We show that (i) both methods provide on par or better disentanglement-reconstruction trade-off without
tuning a β hyperparameter and that (ii) they provide useful insights about the entanglement between the underlying factors.
Third, we introduce a novel correlation-based disentanglement (CD) measure that allows interpreting the degree of disentanglement for each dimension of the representation as well as how well each underlying source of variation is captured by the representation. We demonstrate that CD provides complementary and useful insights
about the disentanglement performance of different DRL methods.