Theoretical Deep Learning #2: Worst-case bounds. Part 2

We continue the discussion of bounding test-train risk difference in worst case. As was noted in the previous part, VC-dimension is unable to explain generalization properties of neural nets. By considering margin loss instead of 0/1 loss, we arrive at different function-class dimension, which we argue to be more suitable for explaining generalization. Find all relevant info on github page: Our open-source framework to develop and deploy conversational assistants: https://