Publications

In our latest work, Shiyu, I and our team explored using LIME, an XAI model local interpretation tool, in explaining radio galaxy morphology classification model rationales. Although detailed analysis is still ongoing, our early result shows that classical CNN (unsurprisingly) would take irrelevant emission regions on the same field image into account when making morphology class predictions.

Common radio galaxy classification could have limitations either when experts with different backgrounds classify on the same image or looking at data of the same source but from different surveys/deep imaging. In this recent work, Micah, me and our team explored the use of the Natural Language Processing (NLP) technique, to derive plain English descriptors for science cases otherwise restricted by obfuscating technical terminology. We generated a set of plain English tags to describe radio source morphology, which will be used in the upcoming Radio Galaxy Zoo:EMU beta test!

In this work, we explored the potential of multi-domain deep learning in finding unusual radio sources such as Giant Radio Galaxies (GRGs). Development of this kind may be valuable in the era of  SKA since SKA and its precursors/pathfinders will have an enormous amount of data await for reliable source finding. We found that multi-domain DL is valuable not only because it did help us get a decent GRG finding algorithm but also allows developers to merge data of different formats when making model predictions, which is somewhat similar to the recent concept of multi-model learning.

​Transfer learning could be one of the practical routines when we are migrating a radio source classification algorithm developed with simulated data to realistic data or trained on one survey but is using it on another survey. Our work several years ago showed that transfer learning can be helpful, though people should also be aware of its limitations.