Objectifs du travail de recherche
Dans ce projet, notre objectif est de corriger cette limitation. Nous chercherons des méthodes pour détecter automatiquement les entrées erronées, corrompues ou invalides. Pour ce faire, nous devrons développer des IA capables d'apprendre l'apparence des images valides, et découvrir comment détecter les images qui ne sont pas conformes à l'apparence apprise. Nous étendrons nos méthodes pour travailler non seulement avec des images individuelles, mais aussi avec des vidéos et des séquences d'images.
Contexte scientifique et social du projet de recherche
Notre travail permettra de développer des IA plus robustes et fiables qui pourront être utilisées en toute sécurité dans des tâches critiques. En combinant la fiabilité de nos méthodes avec la haute performance des IA, nous espérons accroître la confiance en celles-ci et encourager leur utilisation dans des domaines tels que la médecine personnalisée, ce qui, à terme, améliorera la qualité de vie de la société.
Deep learning has lead to huge advances in longstanding problems of computervision, natural language processing, signal processing and robotics. In the nearfuture, deep learning will allow the automation of many critical tasks like medical pathology detection, medical diagnosis, or autonomous driving. However, deep networks are known to provide confident but wrong predictions when they process images that do not belong to the training distribution. Given that acquisition devices, data transmission systems, human operators, and software can all fail in unexpected ways and lead to corrupted images, a deep network might produce inaccurate results, if not, dangerously misleading ones, in these critical tasks.
This project will focus its research efforts in developing a data validation method that can identify whether an input image can be safely processed by a deep network or not.The project is inspired on our preliminary work on deep data validation, where we designed a simple architecture that reaches state-of-the-art performance. However, data validation is a challenging problem, and the capacities of our method are still very limited for practical purposes.
As such, the current project has two goals. The first is to build a strong data validation system that reaches a performance suitable for real-world applications. The second is to incorporate temporal consistency into the validation to deal with sequences of images.To do so, our project proposes (i) to explore deep learning architectures that exploit the multiscale nature of images in order to detect a large range of potential issues; (ii) to explore the performance of methods for one-class classification and integrate them into our system; and (iii) to devise an architecture inspired on recurrent neural networks that can analyze sequences, thus enabling us to incorporate time consistency to the validation process.
By the end of the project, we will have developed a comprehensive methodology on data validation that will have a potential impact not only in the research community, but also in a large number of industries that will be able to build more robust AI systems. We will disseminate our results through publications, conferences, tutorials and websites.
Last updated:04.03.2022
