François CholletVerified account

Also it's Achilles heel. Learns all kinds of nonsense with only remedy being unattainably large datasets (except Google, Apple, etc)

Do you think this applies to multiple task learning too? Is that what you mean by joint feature learning? I need to get this verbage right.

Hmm that may serve that goal but may not be optimal when goal needs to change or alter

Stacked shallow and weak models each feature only speaks about a small aspect or portion of a pattern so it can be relied upon

You're overreaching. This is more a global vs local utility problem. !Coordination game. Wolpert's Wonderful Life Utility as a +better frame

Feeding word embeddings learned by Word2Vec to an RNN is an example for non-end-to-end training. Sometimes it helps to regularize the model.

If it's possible to apply joint feature learning on other systems, what are they?