I want to understand why the hinge loss isn't differentiable. I see that it has a corner which would make it not differentiable there, but then why can we take its subgradients? As far as I know, the $max \left\{0,1-y_it_i \right\}$ have corners too.
Asked
Active
Viewed 2,139 times
1
-
Possible duplicate of this: http://stats.stackexchange.com/questions/4608/gradient-of-hinge-loss – tchakravarty Oct 12 '16 at 16:57