[10] Chelsea Finn - Learning to Learn with Gradients
The Thesis Review
00:00
The Three Properties of a Mammal Like Algorithm
In some ways you can just view mammal also as this as kind of a black box learner. And so we had a theoretical paper that shows that if you have an extremely large neural network then the initialization can basically allow you to represent any learning procedure with just a single gradient step. Yeah I see. So like expressive power measurement, oh sorry consistency and handling ambiguity. You actually showed that just doing these gradient steps is kind of enough for certain models at least. Like you have full expressive power.Yeah. What's your intuition about the types of solutions that it finds because it's essentially saying that like you move in parameter space and then you're kind of assuming that the parameters that you find
Transcript
Play full episode
