James Nightingale:
@Aleksandra, great talk. For both your simple / hard test problems, did the "real" data you applied the CCN on have a diversity of objects that could act as contaminants, or was the dataset mostly known merging objects? If one rolls this out on LSST, is there an expectation for a big drop in purity due to things that 'look' like mergers?


Aleksandra Ciprijanovic:
— Thanks James for this question. In both examples I showed we have merger and non-merger class, where non-merger can be any kind of galaxy. But this is still very simple and in case of new LSST data we would need to isolate galaxies from other objects first if we wanted to use this type of classifier. We can also alway train a more complex model with different types of contaminants. Additionally we would need to retrain a similar model using simulated LSST mocks to try to be similar to that target data.