Can stable and accurate neural networks be computed? - On the barriers of deep learning and Smale's 18th problem

Code related to the paper "Can stable and accurate neural networks be computed? - On the barriers of deep learning and Smale's 18th problem" by M. J. Colbrook, V. Antun and A. C. Hansen.

Overview of the code

The content of this repository has been divided into four directories.

• ellipses: Code used generate all figures involving ellipses.
• inexact_input: Code used to generate Table 1.
• instability: Code used to perform the instability test on AUTOMAP and FIRENET.
• matlab: Matlab code used to produce Figure S3 and Figure S4.

Within each directory you find further details about code dependencies.

Data

To make all the code run seamlessly, you need to download the data and modify the corresponding paths in each of the scripts. The data can be downloaded from storage_firenet and AUTOMAP network weights from here (3.4 Gb).