AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
The Importance of Generalization in Modular Architectures
In the normal like IID train and test condition all these nicely richly structured models don't actually buy you much or anything in the way of performance. The sufficiently big fixed structure models are actually able to do a much better job of answering some of these complicated questions than I certainly would have expected going into this project. In those settings these kinds of explicit modular architectures still I wouldn't say are essential but are generalized much more effectively than than anything else that we know how to build right.