lecture_7

This lecture:

For sufficiently wide nets with one hidden layer GD learns correct predictions on train set in time independent from number of parameters (https://openreview.net/forum?id=S1eK3i09YQ).

Slides; video.

Next lecture announcement:

Information bottleneck method (https://arxiv.org/abs/physics/0004057).

Phases of learning (https://arxiv.org/abs/1703.00810), critics (https://openreview.net/forum?id=ry_WPG-A-&noteId=ry_WPG-A-).

Representation learning, cross-entropy decomposition (https://arxiv.org/abs/1706.01350).