Issue |
A&A
Volume 653, September 2021
|
|
---|---|---|
Article Number | A22 | |
Number of page(s) | 16 | |
Section | Numerical methods and codes | |
DOI | https://doi.org/10.1051/0004-6361/202140369 | |
Published online | 03 September 2021 |
Deep transfer learning for the classification of variable sources
1
Electronics and Telecommunications Research Institute (ETRI), 218 Gajeong-ro, Daejeon, Yuseong-gu, South Korea
e-mail: dwk@etri.re.kr
2
Max Planck Institute for Astronomy (MPIA), Königstuhl 17, 69117 Heidelberg, Germany
Received:
18
January
2021
Accepted:
31
May
2021
Ongoing or upcoming surveys such as Gaia, ZTF, or LSST will observe the light curves of billions or more astronomical sources. This presents new challenges for identifying interesting and important types of variability. Collecting a sufficient amount of labeled data for training is difficult, especially in the early stages of a new survey. Here we develop a single-band light-curve classifier based on deep neural networks and use transfer learning to address the training data paucity problem by conveying knowledge from one data set to another. First we train a neural network on 16 variability features extracted from the light curves of OGLE and EROS-2 variables. We then optimize this model using a small set (e.g., 5%) of periodic variable light curves from the ASAS data set in order to transfer knowledge inferred from OGLE and EROS-2 to a new ASAS classifier. With this we achieve good classification results on ASAS, thereby showing that knowledge can be successfully transferred between data sets. We demonstrate similar transfer learning using HIPPARCOS and ASAS-SN data. We therefore find that it is not necessary to train a neural network from scratch for every new survey; rather, transfer learning can be used, even when only a small set of labeled data is available in the new survey.
Key words: methods: data analysis / stars: variables: general / surveys / techniques: miscellaneous
© ESO 2021
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.