Verifying a toy neural network

Samuel Gélineau will present his AI Safety side project, gelisam.com/parity-bot, which demonstrates that it is possible to verify that a neural network satisfies a safety property. The neural network is trained to follow a safety property, but that's not enough, because we want the model to follow this safety property on all inputs, not only those which have been tested. So we take a range analysis algorithm which is normally used on code, and we adapt it to work on weights.