[theano-users] Why do loss functions aggregate only on the last axis

[Sharapolas]

I am trying to make a custom loss function by looking at the examples at
https://github.com/fchollet/keras/blob/keras-1/keras/objectives.py .

As far as I understand y_pred and y_true in 2D case have shapes of
[n_observations, n_features]. Why is the axis argument -1 everywhere as if
giving the error for each observation separately within a batch? I assume
that -1 means that the average is taken over the lastaxis