# Deep Transfer Learning for Classification of Variable Sources

**Authors:** Dae-Won Kim, Doyeob Yeo, Coryn A. L. Bailer-Jones, Giyoung Lee  
**arXiv:** [2106.00187](https://arxiv.org/abs/2106.00187)  
**Date:** 2021-06-01  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2106.00187v1/

## Key points

- Authors address challenges in observing light-curves of astronomical sources from surveys like Gaia, ZTF, and LSST
- Deep transfer learning approach proposed to identify variability types due to scarcity of labeled training data
- Neural networks used for single-band light-curve classification
- Knowledge transferability leveraged through transfer learning between datasets
- Initial training involves extracting 16 variability features from OGLE and EROS-2 variables
- Incorporating a small set of periodic variable light-curves from ASAS dataset optimizes model for developing a new ASAS classifier
- Transfer learning approach yields good classification performance on ASAS dataset
- Similar techniques applied using Hipparcos and ASAS-SN data, showing effectiveness in transferring knowledge between datasets
- Transfer learning proves effective in achieving accurate classification results with limited labeled data in new surveys

## Summary

In their paper titled "Deep Transfer Learning for Classification of Variable Sources," authors Dae-Won Kim, Doyeob Yeo, Coryn A. L. Bailer-Jones, and Giyoung Lee address the challenges posed by ongoing or upcoming astronomical surveys such as Gaia, ZTF, or LSST in observing light-curves of billions of astronomical sources. The <kw1> deep transfer learning </kw1> approach is proposed to tackle the daunting task of identifying interesting and important types of variability within these vast datasets due to the scarcity of labeled training data, especially in the early stages of a new survey. This involves using <kw4> neural networks </kw4> for single-band light-curve classification and leveraging <kw5> knowledge transferability </kw5> from one dataset to another through transfer learning. The initial training process involves extracting 16 variability features from light-curves of OGLE and EROS-2 variables. Then, incorporating a small set (e.g., 5%) of periodic variable light-curves from the ASAS dataset optimizes this model for developing a new ASAS classifier. The results demonstrate that this <kw1> transfer learning approach </kw1> yields good classification performance on the ASAS dataset, showcasing its effectiveness in transferring knowledge between different datasets. Furthermore, similar techniques are applied using Hipparcos and ASAS-SN data, reinforcing the idea that training a neural network from scratch for every new survey is not necessary. Even with limited labeled data available in a new survey, <kw1> transfer learning </kw1> proves to be an effective strategy for achieving accurate classification results. Overall, this study highlights the potential of <kw1> deep transfer learning </kw1> in addressing data paucity challenges in astronomical surveys and underscores its utility in efficiently leveraging existing knowledge across diverse datasets for improved classification outcomes.