Chelsea Finn on Meta Learning & Model Based Reinforcement Learning
The Gradient: Perspectives on AI
00:00
Roblinet and the Future of Video Prediction Video Modeling
In the Fitvid paper Mohamed was actually able to get one model to overfit on Roblinet. It was a model that was actually larger than his default model. I usually just you don't think of overfitting as good but in this case I think it's great which means that if we give it more data including more diverse data it will probably be able to fit that data reasonably well. Yeah it's exciting and now just full length of paper I see that there's some robotic experiments as well as some other experiments.
Transcript
Play full episode
