Lightweight framework for learning with side information (aka privileged information), based on Theano and Lasagne. For more info on learning with side information see http://arxiv.org/abs/1511.06429
This is half a question about the library and half a general side-information advice question -- apologies if I'm abusing the github issue feature. I have a problem in which I have side-information for a small but significant subset of my training data. If I train my model without this side-information, it doesn't generalize well, but the information needed to make it generalize well for the full dataset is contained in the side-info I have for my small subset.
I'm considering trying to use this library to make use of that information. It seems like it might be easiest if I frame the problem as multi-task learning, with 0 loss, 0 grad-loss for training examples with no corresponding side-information.