Repair of convolutional neural networks using convex optimization: Preliminary experiments

Recent public calls for the development of explainable and verifiable Artificial Intelligence (AI) led to a growing interest in formal verification and repair of machine-learned models. Despite the impressive progress that the learning community has made, models such as deep neural networks remain vulnerable to adversarial attacks, and their sheer size represents a major obstacle to formal analysis and implementation. In this paper, we present our current efforts to tackle repair of deep convolutional neural networks using ideas borrowed from Transfer Learning. Using results obtained on popular MNIST and CIFAR10 datasets, we show that models of deep convolutional neural networks can be transformed into simpler ones preserving their accuracy, and we discuss how formal repair through convex programming techniques could benefit from this process.