I would like to contribute by adding an algorithm performing the Neighborhood component analysis (http://www.cs.toronto.edu/~fritz/absps/nca.pdf)

Basically, it's trying to learn a distance metric by optimizing the leave-one-out performance on the training data. In a first time, I would first only follow the paper, and then try to add different solving methods in the parameters.

Tell me if you would be interested, if have any recommendations for the architecture or if someone else is already doing this.