Tricky Validation Accuracy Plot

I think this can serve as good interview question ;)

You measure loss and accuracy over some validation set:

And you end up with plots like below. What’s going on?

Discussion