an open-source library dedicated to making machine learning models robust against changes in data distribution. Robust models often generalize better and maintain good performance when input data differs from the training set. This robustification is based on a solid mathematical theory and is made tractable by a dual approach and recent advancements in entropic regularization…
A library dedicated to verifying the local stability of developed neural networks. Using advanced techniques such as Linear Relaxation-based Perturbation Analysis (LiRPA), Decomon automatically derives upper and lower bounds for the predictions of a neural network with perturbed inputs. These bounds are calculated using linear relaxations.
Build 1-Lipschitz Neural Networks to guarantee robustness.