Title:	Identifiable Representation Learning
DNr:	Berzelius-2025-91
Project Type:	LiU Berzelius
Principal Investigator:	Ahmet Zahid Balcioglu <ahmetza@chalmers.se>
Affiliation:	Chalmers tekniska högskola
Duration:	2025-04-10 – 2025-11-01
Classification:	10210
Keywords:

Abstract

Representation leaning is an active research area in machine learning focusing on extracting features from a high volume of high-dimensional data in an unsupervised setting. Contrastive learning is a popular method for learning such representations. The quality of such features is usually analysed by testing in downstream tasks. My research focuses on provable or identifiable learning of such representations in a decision-making setting, where the goal is to identify a mapping from context variables to potential outcomes of actions. Applications of my work include healthcare and autonomous driving. This project will be used as continuing my previous work on identifiable latent bandits, which had a healthcare motivation, and exploring applications to autonomous driving. Furthermore, I will also use it for studying state-of-the-art contrastive learning applications and investigating potential research questions. Reference for previous work: https://arxiv.org/pdf/2407.16239