We are going to build a software package for training neural networks. The software package will incorporate basic notions of continuity and smoothness that have been missing from deep learning. As such, we hope to see improvements across several core areas including optimization, numerics, robustness and generalization.
The project has three components that will support and influence each other: