75 - Reinforcement / Imitation Learning in NLP, with Hal Daumé III
NLP Highlights
00:00
The Importance of Headroom in Deep Learning Models
If you're getting 95% accuracy, it's maybe not worth it. The other thing that's sort of changed with the advent of sort of deep learning stuff is that we basically know now that we can overfit anything. But if you overfit your training data, then rolling in with your own policy versus rolling in with the expert are exactly the same thing. This used to be a small issue. Now I think it's a major issue. It adds all sorts of extra overhead that's really difficult.
Transcript
Play full episode
