Strong gravitational lensing is a promising probe of the substructure of dark matter to better understand its underlying nature. Deep learning methods have the potential to accurately identify images containing substructure, and differentiate WIMP particle dark matter from other well-motivated models, including axions and axion-like particles, warm dark matter etc.
Supervised classification can be difficult when the number of known objects of a particular class is very small. This is usually the case for strong gravitational lensing images, where the number of samples from one or more classes are relatively lower than others. Self-supervised learning (SSL) has proven to outperform standard supervised machine learning models, particularly when the number of data labels available for supervision is low. Moreover, SSL can take advantage of very large unlabelled datasets that would be difficult or impossible to label manually and build meaningful representations. To date, only convolutional neural networks (CNNs) have been used with the SSL technique for strong gravitational lensing data. Transformers or hybrid models (Transformers + CNN) promise more robustness for representation learning but have not been addressed by the community. This project will focus on the development of self-supervised learning techniques with Transformers for strong gravitational lensing data. Furthermore, we will also investigate equivariant techniques in the self-supervised learning context for other strong lensing tasks.
Total project length: 175/350 hours.
Python, PyTorch and relevant past experience in Machine Learning.
Please use this link to access the test for this project. The test is due by April 3rd, however please keep in mind that it takes about 1 week to craft a good proposal and proposals need to be submitted via GSoC portal by April 4
Please DO NOT contact mentors directly by email. Instead, please email email@example.com with Project Title and include your CV and test results. The relevant mentors will then get in touch with you.