Deep Learning Techniques Applied on Imbalanced Astronomical Datasets

Not scheduled
2m
Poster Posters

Speaker

Dr Emmanuel Bratsolis (University of West Attica)

Description

Accurate redshift estimation is a fundamental task in astronomy, underpinning studies of galaxy distances and large-scale structure, and providing key constraints on galaxy evolution. In this work, we investigate the application of deep learning techniques to predict galaxy redshifts from spectroscopic data. We develop and train a neural network model using galaxy spectra obtained from the European Space Agency’s Gaia mission.
Given the highly imbalanced nature of the available dataset—particularly at high redshift—we explored a range of data augmentation strategies, loss functions, and alternative model architectures to mitigate this issue. These investigations informed the design and training of our final model. The results demonstrate strong predictive performance overall.
In particular, a comparison between our Hybrid Convolutional Neural Network–Transformer (HCNN-Transformer) model and a standard convolutional neural network (CNN) shows that the proposed approach represents a significant improvement over the CNN baseline. The HCNN-Transformer more effectively captures high-redshift galaxies, a regime in which the simpler CNN model performs poorly. Further performance gains appear possible through the use of advanced sampling techniques, although these may come at increased computational cost. Notably, the HCNN-Transformer model also achieves a substantial improvement over the redshift estimation performance reported for Gaia DR3.
Beyond astronomy, the methods explored here are broadly relevant to scientific domains that must address learning from strongly imbalanced datasets.

Author

Dr Emmanuel Bratsolis (University of West Attica)

Co-authors

Mr Konstantinos Peratinos (University of West Attica) Prof. Despina Hatzidimitriou (National and Kapodistrian University of Athens) Prof. Ioannis Bellas - Velidis (National Observatory of Athens)

Presentation materials

There are no materials yet.