[24] Martin Arjovsky - Out of Distribution Generalization in Machine Learning
The Thesis Review
00:00
Neural Neural Networks - Is There a Path to Generalization?
In your thesis you're taking this approach that's based on the objective right so it's like developing what is the right objective to yield something which has these properties we want. Do you think that we also need to think about or would it be like a separate path where we think about the actual parameterization of the model? I'm very alone in my opinion by the way but both are going to be needed, he says. Even if you have the perfect loss unless you have a model that thinks algebraically you're not going to learn algorithms or real causality as well.
Transcript
Play full episode
