Transfer learning for transient search with small-field optical survey telescopes

Technical Summary: Transfer Learning for Transient Search with Small-Field Optical Survey Telescopes

Problem Statement
The advent of automated optical sky surveys has necessitated the use of supervised machine learning (ML) to efficiently search for transient events. While Convolutional Neural Networks (CNNs) have proven effective for distinguishing real transients from artifacts (bogus detections) and classifying them, their successful deployment typically requires extensive, manually labeled training datasets. Small-field survey telescopes, such as the 4-m International Liquid Mirror Telescope (ILMT), face a significant challenge: they lack the volume of labeled data required to train robust deep learning models from scratch. Training on insufficient datasets risks overfitting, where models perform well on training data but fail to generalize to unseen test data. Furthermore, existing ILMT frameworks rely on synthetic datasets, which may introduce systematic biases and fail to capture the complexity of real-world transient events.

Methodology
This study proposes the application of Transfer Learning (TL) to overcome data scarcity in the ILMT transient search pipeline. The methodology involves a two-step process:

1. Source Model Training: A CNN-based classifier is first trained on a large, labeled "source" dataset derived from the Zwicky Transient Facility (ZTF). This dataset includes image cutouts (triplets of science, reference, and difference stamps) labeled by the ALeRCE broker. Three distinct source models were developed:
   • A binary real/bogus classifier.
   • A 3-class transient alert classifier (Extended-host, Point-source, Orphan).
   • A 4-class transient alert classifier (Extended-host, Point-source, Orphan, and Bogus).
2. Target Model Adaptation: The pre-trained source models are adapted to the "target" domain (ILMT data) using fine-tuning. To address the difference in pixel scales between ZTF (1 arcsec/pixel) and ILMT (0.323 arcsec/pixel), ILMT stamps were downsampled to match the ZTF scale.
   • For the real/bogus classifier, the entire network was fine-tuned on the smaller ILMT dataset with a reduced learning rate ($10^{-5}$) to preserve general features learned from ZTF while adapting to ILMT-specific characteristics.
   • For the 3-class classifier, a "layer freezing" approach was employed where the convolutional base was frozen, and only the fully connected classifier layers were trained.
   • For the 4-class classifier, the network was fine-tuned with a reduced learning rate ($5 \times 10^{-5}$).

The target datasets were constructed from ILMT observation cycles (October 2023–May 2024 for training; October 2022–November 2022 for testing) and manually vetted. Data augmentation (rotations) was applied to under-represented classes in the target set.

Key Contributions

• Proof of Concept for TL in Small Surveys: The paper demonstrates that knowledge distilled from large-scale surveys (ZTF) can be effectively transferred to small-field facilities (ILMT) to train CNNs for transient detection and classification, bypassing the need for massive local labeled datasets.
• Dual-Classifier Framework: The study introduces and validates two types of alert classifiers: a 3-class model for classifying confirmed "real" candidates and a 4-class model that integrates artifact rejection directly into the classification stage.
• Operational Integration: The trained TL models were successfully integrated into the existing PyLMT pipeline and deployed on full-frame ILMT images, moving beyond theoretical validation to operational application.
• Statistical Validation: The effectiveness of the TL approach was rigorously tested against baseline models (trained without TL) using an unpaired t-test, confirming statistically significant improvements in performance.

Results
The models were evaluated on independent test datasets derived from earlier ILMT observation cycles.

• Real/Bogus Classification: The TL-based model achieved a test accuracy of 97.3%, significantly outperforming the baseline model (trained only on ILMT data without TL) which achieved 90.5%, and the source ZTF model applied directly to ILMT data (91.5%).
• Transient Alert Classification:
  • The 3-class classifier achieved 92.9% accuracy with TL, compared to 77.3% for the baseline.
  • The 4-class classifier achieved 85.6% accuracy with TL, compared to 75.5% for the baseline.
• Statistical Significance: Unpaired t-tests confirmed that the mean accuracies of the TL models were significantly higher than the baseline models across all three classifier types (p-values $< 0.05$).
• Operational Performance: When deployed on ~300 full-frame ILMT images, the pipeline successfully detected various astrophysical phenomena, including asteroids (median V-band magnitude ~20), variable stars, AGNs, and supernova candidates (e.g., AT 2024abso, AT 2025chp).

Significance and Claims
The authors position this work as a proof of concept for enhancing transient detection in small-field surveys, rather than an immediate replacement for existing schemes. The paper claims that the fundamental image-level signatures of genuine transients and subtraction artifacts are largely transferable across optical surveys when domain differences (such as pixel scale) are mitigated.

The primary significance lies in the ability of small surveys to leverage knowledge from large-scale facilities to accelerate the deployment of reliable transient pipelines without waiting years to accumulate sufficient labeled data. The authors note that even modest gains in real/bogus discrimination can substantially reduce the manual vetting burden, which is critical for small surveys with limited human resources. However, the paper modestly acknowledges limitations, including the relatively small size of the target dataset, the partial reliance on injected sources for certain subclasses, and the potential for data drift due to evolving observing conditions, suggesting that future work should explore continual learning frameworks.