On the Connection between Neural Networks and Kernels: a Modern Perspective -Simon Du 

It feels intuitive that global average pooling should have some value... like to Hinton's capsule networks, except the capsules are blasted to pieces and diluted into an ensemble, rather than discrete.

at 4:55 there is a promise of complexity measure that explains what VC-dim cannot. At the end of the talk, I fail to see what that measure is.

at 10:40, under Trajectory of predictions, the sum of angle brackets -- there are two scalars inside scalar brackets???

Why are the presentation slides not provided? This talk is very math heavy

Amazing work, but sorry... I cannot understand the speaker's English most of the time. The very frequent mismatch between the RU-vid automatic subtitles and the speaker's words corroborates my opinion.