USING VARIATIONAL AUTOENCODERS TO ENCODE GALAXY IMAGES IN SEARCH OF NEW GRAVITATIONAL LENS CANDIDATES

• Main Author: L. G. Tampubolon, Feralezer
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/76656
• Institution: Institut Teknologi Bandung

Discovering space phenomena such as gravitational lensing among the vast array of space images available today can be a daunting task if done manually. However, it can be made more manageable with the right tools, such as a deep learning-based astronomical image retrieval subsystem developed in this final project. To function effectively, this subsystem requires a deep learning component capable of encoding astronomical images into vector representations which are retrievable. Consequently, this project also involves training a type of deep learning model called variational autoencoder, whose encoder can transform images into vectors. Our variational autoencoders are trained using a dataset consisting of 88,422 galaxy images, with 40,500 of them being gravitational lens images. Upon testing against a validation set, our deep learning-based astronomical image retrieval subsystem achieves an NDCG score of 0.722. Further testing on unseen and end-to-end data respectively yields 0.603 and 0.574 NDCG scores. These outcomes indicate that our deep learning-based astronomical image retrieval subsystem shows tremendous promise and can be integrated into a gravitational lens candidate search system based on deep learning principles.