FINDING NEW GRAVITATIONAL LENS SYSTEMS USING NOISY STUDENT TRAINING

• Main Author: Alexander Suryasumirat, Naufal
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/75267
• Institution: Institut Teknologi Bandung

The search for new gravitational lens systems is still a big problem faced in the field of astronomy. Besides being able to be used to prove the theory of general relativity, the phenomenon of gravitational lensing can also be used to observe objects at great distances that are usually cannot be observed from Earth. Until recently, the most commonly used approach to find new gravitational lenses is supervised deep learning. However, there are certain drawbacks to this approach, namely its poor performance for classifying data that the model has not seen before. Therefore, a semi-supervised approach is used in the form of Noisy Student Training (NST) that is both iterative and expansive to try to solve this problem. This semi-supervised approach has several key advantages such as it has been proven to increase the performance of state-of-the-art models for ImageNet classification problems and also requires a small amount of labeled data. In this final project, the implementation of NST and the new gravitational lens systems finder model generated by the NST process have been successfully developed. The best model from all NST iterations has better performance than the teacher model in the initial iteration and shows the capability to find new gravitational lens systems. The evaluation was carried out on the models resulted from each NST iteration based on the metrics of the model to classify the new gravitational lens systems such as accuracy, precision, recall, and f1-score. However, based on the results of the evaluations done in this final project, the performance improvement is considered less effective and less efficient than ideal. Only one model shows performance improvement in classifying new gravitational lens systems out of every NST experiments done. Furthermore, the performance produced by the model at each NST iteration can be observed as fluctuating. The method is inconsistent in producing a better model than the initial model for the dataset used. This is also caused by the stochastic nature of the NST approach.