verifiNN 0.0.0.dev1

A package for optimization based neural network verification.

verifiNN

This project implements the training of deep linear networks from the ground up. We also intend to use it in investigating the loss landscape of deep linear networks and comparing it with other learning models, e.g. linear rigression. We hope to publish the results of our learing in the forrm of blog posts.

Prerequisites:

• python 3
• virtualenv
• pip

Instructions:

• set up the python virtual enviroment by running setup.sh
• add the following entry to the ~/.bash_profile file export PYTHONPATH="<path/to/parent/repo>:$PYTHONPATH"
• start up the python environment by . ./venv/bin/activate

References:

• https://arxiv.org/pdf/1811.01057.pdf (main paper)
• https://arxiv.org/pdf/1903.06758.pdf
• http://www.cs.cmu.edu/~cliu6/16-883/Lecture%2010%20-%20Optimization-Based%20Verification%20of%20Deep%20Neural%20Networks.pdf