Christoph Sawade

Add to list: I guess also the noncentral-hypergeometric distributions are missing (http://en.wikipedia.org/wiki/Noncentral_hypergeometric_distribution).

I am thinking about a confidence interval for the sign test. Currently, there are actually two sign tests, one for testing the median of a distribution and a paired one...

Well, I mean Pegasos is only an instance of a stochastic gradient descent method (aka Robbins-Monro algorithm). This is applicable for multiple objectives that can be seen as an (empirical)...

Yes. But I have to think about it a bit more careful - hopefully next weekend.

Hi a put my attempt online. You can find it here: https://github.com/BigCrunsh/RegERMs.jl.

@jumutc: Does it make sense to consider this in an even more general framework, too. The work on the regularized empirical risk framework (https://github.com/JuliaStats/RegERMs.jl) has been dormant for some time....

@jumutc: yes definitely. Here some comments: 1) yes the `optimize` method is currently specific to SGD and L-BFGS. However, I think it makes sense to generalize or extend it to...

I agree with @johnmyleswhite, unification is an ongoing and expensive process. However, that shouldn't stop you / us tackling the points you (@jumutc) mentioned. Specifically, 1) yes, long-term this makes...

I agree in principal with these abstract types. For the loss, however, I suggest to have a function that takes a decision value and a label, since there are differences...

or just a different signature `value(l::LogisticLoss, f::Real, y::Real)` and `value(l::LogisticLoss, f::Vector, y::Real)`?