Causal and identifiable deep representation learning

Thesis

Deep Learning has been responsible for many of the recent artificial intelligence (AI) successes, from text generation, image understanding, protein structure prediction, to superhuman performance in playing chess. Despite their successes, most current deep learning methods have difficulties with generalisation beyond the training distribution, are often not interpretable or explainable, possess few learning guarantees, and are often not robust to environmental changes. As AI is being adopted in more and more parts of society, it is important to understand and address these limitations.

One way to address these issues is by applying the framework of causality to deep learning, in what has become the emerging field of causal representation learning. In causal learning, the emphasis is on using data to learn causal variables and relationships instead of pure statistical associations, as this allows for improved interpretability, transfer, and reasoning via the framework of do-calculus.

In this thesis we take inspiration from causality and identifiability to develop deep learning methods for visual data that are interpretable, possess learning guarantees, and generalise beyond the training distribution. We do this by designing deep learning systems that employ various assumptions, such as assumptions about the environment (e.g. video with a static background), assumptions about the latent space (e.g. having one or two variables, or using a causal graph), restrictions on the causal mechanism function classes (e.g. linear functions), and properties of the network architectures (e.g. equivariance to certain transformations). Using these assumptions allows us to prove that the models are guaranteed to learn the true latent variables (up to some transformations), show that the models successfully generalise beyond the training distribution by intervening on the latent representation and generating realistic videos never observed at training time, and demonstrate that the learned representation is easily interpretable and explainable.

Authors: Longa, M

• DOI: 10.5287/ora-nrjz0pexj
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: machine learning; deep learning; computer vision; representation learning; causality; artificial intelligence; equivariance
• Subjects: Artificial Intelligence; Deep Learning; Machine Learning; Computer Vision