ON DISENTANGLED REPRESENTATION LEARNING

Date

2024-7

Author

MOĞULTAY ÖZCAN, HAZAL

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

84
views

135
downloads

Cite This

Disentanglement 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 introduces a new hyperparameter β to weight the disentanglement term of the VAE loss function, which is empirically optimized. In this thesis, we make three contributions to improve the disentanglement capacity of β-VAE. First, to automatically estimate the β parameter, we propose Learnable VAE (L-VAE). L-VAE, mitigates the hyperparameter optimization problem of β-VAE by learning the relative weights of different terms in the loss function to dynamically control the trade-off between disentanglement and reconstruction. These weights and the parameters of the model architecture are learned concurrently, eliminating the complexity of empirical hyperparameter optimization. Both β-VAE and L-VAE introduce the same weight on all of the dimensions of the representation. However, we show that each dimension has a different degree of disentanglement. In order to dynamically learn an independent weight per dimension, we propose Multi-Dimensional Learnable VAE (mdL-VAE) as an extension to L-VAE. We show that both methods provide on par or better disentanglement-reconstruction trade-off without tuning a β hyperparameter and that mdL-VAE provide useful insights about the entanglement between the underlying factors of variations. Finally, we introduce a novel correlation-based disentanglement (CbD) measure that allows interpreting the degree of disentanglement for each dimension of the representation robustly. We demonstrate that CbD provides complementary and useful insights about the disentanglement performance of different disentanglement methods.

Subject Keywords

Disentanglement, Representation learning, Disentanglement measure, Hyperparameter optimization

URI

https://hdl.handle.net/11511/110813

Collections

Graduate School of Natural and Applied Sciences, Thesis