Do Neural Networks Dream of Gravitational Lenses: Using CNN to Identify Gravitational Lenses & How They Do It
AstroML
Josh
Wilde
Date Submitted
2021-04-30 00:00:00
Open University
J. Wilde (Open University), S. Serjeant (Open University), J. Bromley (Open University), H. Dickinson (Open University)
In preparation for future large surveys such as LSST and Euclid. These expect to find more than 10^5 gravitational lenses, I am interested in making sure the novel systems are discovered as these offer greater constraints on dark matter. I have been developing a CNN model to identify gravitational lenses from simulated Euclid images. This CNN model performs well with an F1 score of 0.98, but why? I have applied several approaches including deep dream, occlusion maps, and class generated images to understand the aspects of the image which influences the model’s classification. Currently I am creating images of compound lenses to understand how well my model performs on data of rare lens configurations.
All attendees are expected to show respect and courtesy to other attendees and staff, and to adhere to the NAM Code of Conduct.