Abstract The advent of optical sky surveys has enabled several automated programs for searching transients. Many of these programs extensively use supervised machine learning (ML) algorithms to automate these searches. Effective implementation of such a strategy has an advantage over non-automated methods of transient search in terms of reduced manual labour and reporting latency. Training the relevant ML algorithms often requires extensive labelled training datasets that might not be readily available for new or small field-of-view survey telescopes. Transfer Learning (TL) is an ML technique that is often employed to address this issue by transferring knowledge from a pre-trained model, trained on an extensive dataset for a related task, to enhance performance on a new task with a limited dataset available. This paper demonstrates TL for a Convolutional Neural Network (CNN)-based real/bogus classifier model for transient detection between extensive publicly available image data from the Zwicky Transient Facility (ZTF) and a small and labelled dataset from the 4-m International Liquid Mirror Telescope (ILMT). The same technique was employed to train two different types of transient alert classifiers to characterise the detected candidates based on detection image stamps into 3 and 4 classes, respectively. The resulting model for the real/bogus classifier achieved an accuracy of 97.3% on the test dataset. Additionally, an accuracy of 92.9% was achieved for the 3-class classifier and 85.6% for the 4-class classifier. Furthermore, the statistical significance of the effectiveness of this technique was established with an unpaired t-test between TL models and baseline models trained without TL.
Pranshu et al. (Tue,) studied this question.