Hi there,
I’m probably late…did you find a solution? I’m strugging with a similar problem. At some point in the beginning of the course (2018 version) someone asked Jeremy that and he said that a common approach is to oversample the less frequent classes, i.e., you duplicate some training examples (he also said, that we’d be talking more about that in future lessons). I’m now trying that but still having problems (validation loss goes up whereas the training loss goes down => overfitting).
Cheers.