search

uchicago-computation-workshop / nicolas_masse

Repository for Nicolas Masse's presentation at the CSS Workshop (1/13/2019)
0 stars 0 forks source link

GAN’s + XdG #5

Open AlexanderTyan opened 5 years ago

AlexanderTyan commented 5 years ago

In regards to a model that can distinguish between contexts, switch and perform necessary gating, do you think Generative Adversarial Networks could be of much use? (This is in the context of progress toward transfer leaning). As to a non-expert in ML, it would seem to me that the idea of competing networks (like GAN’s) might be of application when a model decides which “context” and gating is best for a particular task.

Sent with GitHawk

nmasse commented 5 years ago

Determining what context the agent is in a principled manner is very high on my wish list! It's a tough problem because, frankly, neuroscientists aren't even quite sure what "context" actually means. However, it is widely believed that the hippocampus and related areas are essential in binding what we learn/remember with the context we're in, and we're currently working on ideas along these lines. Using GANs on the other is an interesting idea, and warrants some thought. However, they can be real difficult to train, and would have to think carefully before proceeding.