Deprecate Linear builder?

[ablaom]

The MLP builder added more recently renders it redundant.

[EssamWisam]

Agree. Think should also deprecate Short because it's also a special case of MLP and forces specifically three hidden layers (also quite hard to infer the functionality from the name; a 3 layer network can theoretically even approximate discontinuous functions to arbitrary accuracy).

[ablaom]

'Short' includes a dropout layer and for this reason is needed in tests, if I remember correctly. I agree the name is not great and we could deprecate to "testing only".

[EssamWisam]

Sheds light on the possibility of extending MLPs to handle dropout (and maybe batch norm because it's just as common). For instance, for drop out, user can pass a list of probabilities that correspond to the hidden layers and 0 would mean no dropout.