Causal disentanglement aims to learn representations of causally related latent variables, wherein the data pertaining to the variables of interest are observable only through an unknown linear transformation. This project studies how we can perform causal disentanglement from the score of the observed data distribution in non-linear additive Guassian noise models without interventional data. Both general methods for causal disentanglement and the use of score-matching for causal discovery have been at the forefront of recent causal inference research, yet a method of using score-matching for causal disentanglement without strong assumptions on the presence of interventional data has yet to be discovered. The algorithm developed in this project will be the first of its kind and can be applied to learning casual representations with latent variables in numerous settings such as economics and computational biology.

I am participating in SuperUROP to gain experience in performing theoretical research. All of my prior research experiences have been very high level, and I have always been interested in diving deeper into the the theoretical models that drive many machine learning applications. I hope to contribute significantly to the field of causal representation learning this year and pave the way for a successful future in this line of research.