aka multi-instance learning, multi-task learning

See Deep learning

Max-margin learning, transfer and memory networks.

good for generalizing models, . Good when don't have much supervision data.

Max-margin learning

Learn embeddings in one task and transfer these to solve new tasks

Example. He exaplains how deep multi-instance learning works. Nice

Matching

Corruption (and example here)

Example: Bi-lingual word embeddings

When you can't corrupt the data: Siamese networks Paper

Example: Question answering system. Followed by relation learning (learning triplets like "cat eats mouse")

memory networks (see below) may be useful for transfer learning too..

One-shot learning using conv nets, as we've already have good embeddings, just compare objects in embeddings. See beginning of this

See also Feature selection

See also Incremental learning