Open zenogantner opened 11 years ago
Those different methods all support optimizing for MAE (not the default!), but we need to make sure that the learning algorithm (sub-gradient descent) always behaves nicely, in particular in those areas in which MAE is not really differentiable.
First step: mark the functionality as unstable.
Those different methods all support optimizing for MAE (not the default!), but we need to make sure that the learning algorithm (sub-gradient descent) always behaves nicely, in particular in those areas in which MAE is not really differentiable.