The promise of self-supervised and active learning for Strong Lens discovery: Astronomaly applied to KiDS
Abstract
Strong gravitational lenses (SGLs) are rare systems whose discovery currently relies primarily on supervised machine learning methods trained on large simulated datasets. We present the first application of Astronomaly:PROTEGE to SGL discovery, demonstrating that a human-in-the-loop active learning framework can efficiently identify lenses in large imaging surveys without the need for simulated training data. We consider a sample of 3.7 million bright galaxies from the Kilo-Degree Survey (KiDS) DR4. Feature representations are extracted using a convolutional neural network pre-trained on the ImageNet dataset and subsequently fine-tuned on KiDS data using the self-supervised Bootstrap Your Own Latent (BYOL) framework. Within the embedding of these representations, the active learning loop of Astronomaly iteratively selects the most informative systems for expert inspection. A total of 3,000 objects are inspected across multiple rounds, yielding 34 high-quality (grade A/B) SGL candidates. On the basis that these systems occupy similar regions in the learned feature space, we expand this sample through nearest-neighbour similarity analysis. Including the active learning discoveries, we identify a total of 140 grade A/B candidates and more than 1,000 additional lower-confidence systems (grade C). Among the A/B candidates, 81 are newly identified, while approximately 22% of previously known grade A/B KiDS lenses are recovered. These results demonstrate strong potential for next-generation surveys such as Euclid, Roman, and Rubin's Legacy Survey of Space and Time. With approximately 60% of the high-quality candidates newly reported, this approach complements supervised methods by reducing reliance on simulations and enabling the discovery of a diverse population of SGLs.
Related
- The NEMESIS general young stellar object catalogue. I. Supervised classification with deep learning methods
- Self-Supervised Representation Learning for JWST Galaxy Images and Low Resolution Spectra
- Lyman Break Galaxy selection and redshift measurement with supervised contrastive learning
- Self-Supervised Representation Learning: From Spectral Foundation Models to Auroral Emission Spectra
- A self-supervised machine learning approach to identify instrumental systematics in 21cm Cosmology
BibTeX
@misc{grespan2026promise,
title = {{The promise of self-supervised and active learning for Strong Lens discovery: Astronomaly applied to KiDS}},
author = {Margherita Grespan and Aprajita Verma and Michelle Lochner and Koketso Mohale and Verlon Etsebeth and Duncan Bowden},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.00154},
eprint = {2609.00154},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.00154},
}
RIS
TY - UNPB
TI - The promise of self-supervised and active learning for Strong Lens discovery: Astronomaly applied to KiDS
AU - Margherita Grespan
AU - Aprajita Verma
AU - Michelle Lochner
AU - Koketso Mohale
AU - Verlon Etsebeth
AU - Duncan Bowden
PY - 2026
DA - 2026/09/03
PB - arXiv
DO - 10.48550/arxiv.2609.00154
UR - https://arxiv.org/abs/2609.00154
AB - Strong gravitational lenses (SGLs) are rare systems whose discovery currently relies primarily on supervised machine learning methods trained on large simulated datasets. We present the first application of Astronomaly:PROTEGE to SGL discovery, demonstrating that a human-in-the-loop active learning framework can efficiently identify lenses in large imaging surveys without the need for simulated training data. We consider a sample of 3.7 million bright galaxies from the Kilo-Degree Survey (KiDS) DR4. Feature representations are extracted using a convolutional neural network pre-trained on the ImageNet dataset and subsequently fine-tuned on KiDS data using the self-supervised Bootstrap Your Own Latent (BYOL) framework. Within the embedding of these representations, the active learning loop of Astronomaly iteratively selects the most informative systems for expert inspection. A total of 3,000 objects are inspected across multiple rounds, yielding 34 high-quality (grade A/B) SGL candidates. On the basis that these systems occupy similar regions in the learned feature space, we expand this sample through nearest-neighbour similarity analysis. Including the active learning discoveries, we identify a total of 140 grade A/B candidates and more than 1,000 additional lower-confidence systems (grade C). Among the A/B candidates, 81 are newly identified, while approximately 22% of previously known grade A/B KiDS lenses are recovered. These results demonstrate strong potential for next-generation surveys such as Euclid, Roman, and Rubin's Legacy Survey of Space and Time. With approximately 60% of the high-quality candidates newly reported, this approach complements supervised methods by reducing reliance on simulations and enabling the discovery of a diverse population of SGLs.
ER -