Deep Learning Foundations: Misha Belkin's Talk on deep learning through the prism of interpolation

[@yaram.bahram6715]

Very interesting remarks on generalization. Thank you for sharing this great lecture!

[@jfjfcjcjchcjcjcj9947]

Thanks for the lecture! I think that there are a couple of points that would make the lecture more digestible if they are explained a bit more in depth. First, is the definition of interpolation, different people imply different things when referring to interpolation. I presume that Misha here by interpolation refers to the ability of estimators to achieve zero training error and still not overfit to the hold-out test set? The second aspect which might require some more explanation is @40:05 when Misha is describing the difference between interpolation accuracy and interpolation in the l2-norm. I presume that what is meant by l2-norm interpolation is that 2 estimators with the same l2-norm on their parameters exhibit the same generalisation capabilities, is that right?

[@alexisfernandez8052]

I can show you how! Let know!

[@TheMaP142]

Michael