What can linearized neural networks actually say about generalization?

Open in new window