Lavanya Nemani 02:58 PM @Aritra Thanks for the very interesting talk! Could you elaborate more on the transformation you make on images of different sizes to make them of the same size? Aritra Ghosh 03:04 PM Thank you Lavanya. So, essentially we are using something which is known as a spatial transformation network. This STN itself contains a small neural network which predicts a helpful transforamtion to make (so, essentially what this small neural network is predicting are the six parameters of an affine transformation). I hope this clarifies a little bit more how we are doing this. However, if it’s not, please don’t hesitate to send me an e-mail. Fionagh Thomson 02:59 PM what do you see as the potential medical applications? Aritra Ghosh 03:09 PM I think I would see this general framework being used for any medical imaging diagnosis work, where there is not too much real data available to train a deep learning algorithm. (although I will admit that I don’t know too much about simulating medical imaging data) and also, similarly, the interpretability techniques that we use can be helpful for any deep learning algorithm which are being used on medical imaging data; where I guess interpretability is even more crucial. Iary Davidzon 3:01 Very interesting! I was wondering how you added realistic noise to the simulation (I feel this is always the toughest step in the “domain transfer” procedure) Aritra Ghosh 03:02 PM Hi Iary, Thank you. So, what we do is for our simulations is that we add actual real noise from the data that we want to classify. We take large swathes from different parts of survey, mask out the sources, and then read in all the pixels that are leftover as noise. Alternatively, we have also tried adding noise using GalSim’s noise module which also does a decent job. It’s important to note here that since we are doing the final transfer learning step on actual real data (where the noise is also real inherently), even if the noise in the simulations is not 100% accurate, the transfer learning step makes the network learn the noise as well.