Graph-based methods coupled with specific distributional distances for exploration of artificial neural networks

Artificial neural networks are for example prone to being fooled by carefully perturbed inputs which cause an egregious misclassification. Graph theory has been extensively used to model real data such as brain, social interactions and others. In this talk, I will show how graphs-based approaches help to investigate the inner workings of artificial neural networks in two different experiments: adversarial attacks and catastrophic forgetting.