A Stanford Intelligent Systems Laboratory (SISL) research group has announced it is open-sourcing its NeuralVerification.jl project, which helps verify deep neural networks’ training, robustness and safety results.
The library is now available in GitHub and contains implementations of various methods used to verify deep neural networks. The resource divides methods to verify whether a neural network satisfies certain input-output constraints into five categories, including:
- Reachability methods: ExactReach, MaxSens, Ai2,
- Primal optimization methods: NSVerify, MIPVerify, ILP
- Dual optimization methods: Duality, ConvDual, Certify
- Search and reachability methods: ReluVal, DLV, FastLin, FastLip
- Search and optimization methods: Sherlock, BaB, Planet, Reluplex
The library’s installation instructions are as follows:
Publication of the Stanford team’s related research paper is expected by the end of January 2019 at the earliest.
Author: Victor Lu | Editor: Michael Sarazen