Toward scalable and unbiased scene graph generation: Active learning and causal inference perspectives

This dissertation addresses machine learning methods used to generate visual structures based on data perceived by computers. Specifically, the study examines two major challenges: how to reduce manual annotations and how to improve the predictability of systems by mitigating prediction biases. This dissertation introduces a novel approach to solve these problems by combining active learning with causal inference.

The study presents a model named EDAL, which reduces the amount of labeled data required and improves predictability with very small data sets. The second part of the research explores how biases in predictions, such as long-tail distributions and semantic errors, can be eliminated through causal methods. This research introduces three causal approaches that enhance prediction accuracy and help create fair systems.