Mechanistic interpretability for astronomical/cosmological image models

Project Description: 

We propose a project focused on mechanistic interpretability for astronomical/cosmological image models. The project will involve training neural networks on an astronomical/cosmological imaging dataset and applying the ViT-Prisma framework to analyze the learned representations. The goal is to investigate how internal components of the model encode astrophysically/cosmologically meaningful features and to extract interpretable insights about what the network learns from astronomical/cosmological images. This project will introduce the student to modern deep learning techniques for astronomy/cosmological while exploring emerging approaches in mechanistic interpretability. Prior knowledge of machine learning is helpful but not required for highly motivated students.(UCT)

Supervisor

University of Cape Town (UCT)

Co-Supervisor

UCT