DEEP-VISION – DEEP-VISION
Chair of research and teaching DEEP-VISION
Towards Frugal, Trustable, Explainable and Privacy-Aware Deep Learning for Computer Vision
Chair Description
The work program of the Chair will be conducted in 3 directions, all applied to the domain of Computer Vision : i) privacy-aware federated learning ii) explainable deep learning iii) frugal and efficient deep learning algorithms.<br /><br />Towards Privacy-Aware Federated Learning: The introduction of machine learning tools, in a society now massively built around the exchange of digital information, raises growing concerns about protection of privacy. This is all the more true for deep learning – with models having incredible capacity to memorize large volume of information – which has revolutionized machine learning with unprecedented performance, e.g., in face recognition. To address the privacy threats on personal data such as face pictures [O-8], lawyers and experts from many countries more and more emphasize on the principle of “privacy by design”2. This proposal is concerned by the question of privacy aware federative learning. Indeed, a privileged way to build large databases necessary to drive algorithms is to aggregate different sources of content (e.g. medical images held in several hospitals). But, in such situations, how to use the data for training algorithms while preserving privacy is still an open question. One key idea is that training must be done in a federated way without exchanging any private information. Under such constraints, is it even possible to train deep algorithms?<br /><br />Towards Explainable Deep Learning: In many industrial fields, services or highly regulated areas such as banking or aerospace, it is crucial to have efficient systems. However, applying them in production requires an understanding of how to get the results, in order to help their relevance and understand their applicability. Cedric Villani, in his report on AI, calls this ”open the black box”: The explainability of learning-based systems is a real scientific challenge, which puts in tension our need for explanation and our concern to efficiency.<br /><br />Towards More Frugal and Efficient Deep Learning Algorithms: The construction of efficient neural networks is tending towards increasingly deeper networks with ever more parameters. These networks are then difficult to learn, require more and more learning data, and also require a large number of computing resources to be used. Nevertheless, the principle of the razor Ockham states that it is generally preferable, to obtain a good capacity of generalization, to consider solutions with the least possible parameters and knowledge transfer from available sources of information. Statistical learning is generally a search for a good tradeoff between a large number of parameters making it possible to define complex decision boundaries and a smaller number allowing good generalization properties to unknown data. Our overall goal is to develop computer vision algorithms requiring less training data, less computational resources without making concessions on the performances.
Towards Privacy-Aware Federated Learning: The methodology adopted in the following is based on two pillars: i) the use of differential privacy [O-3] and, ii) the introduction of artificially generated data used as proxy for real data. Regarding the first point, we plan to build on the very recent work of [O-6]. The originality will be on the second point, with the proposition of novel methods for generating artificial image with Generative Adversarial Networks [O- 5]. We already made a first step in [42] and plan to continue in this direction, relying on the idea that generative networks may be robust against membership attacks. We want to offer a general-purpose solution to the membership privacy problem by proposing data construction methodologies to produce such surrogate datasets. Our first step will be to use images given by GAN generators, labelled with the classifiers trained on a private dataset. Indeed, we showed that such surrogate data can further be used for a variety of downstream tasks such as classification and regression, while being resistant to membership attacks.
Towards Explainable Deep Learning: We have already made several contributions to this field by proposing image representations that are semantic by design, as in [3]. Here, we will go one step further by addressing the question of the evolution of systems providing semantic description of images over time, through an incremental approach. This will be the topic of the PhD. The second research direction (postdoctoral fellow) will address the question of deep learning based image editing finely controllable by users, through the use of meaningful explainable parameters.
Towards More Frugal and Efficient Deep Learning Algorithms: We will focus our efforts on two particular tasks: on the one hand, the detection of objects in aerial images, a task for which obtaining large databases is difficult. On the other hand, we will be interested in Vision for Robotics tasks where it is important to have methods that guarantee response times adapted to each situation (sometimes it is better to have a less precise inference, but more quickly). Regarding the first task, we have proposed in [2] an original direction, which we would like to deepen. Regarding the second direction, we have also made some interesting proposals in [29].
This section will be enriched as the project progresses.
This section will be enriched as the project progresses.
This section will be enriched as the project progresses.
Ce projet de Chaire IA est porté par Frédéric Jurie. Frédéric Jurie est professeur des universités (PRCE) à l'Université de Caen Normandie, après avoir été chercheur au CNRS de 1994 à 2004 à l'UMR LASMEA à Clermont-Ferrand, puis de 2004 à 2007 au centre INRIA Rhône-Alpes, affilié au groupe de recherche Lear de Cordelia Schmid.
Le travail de recherche du candidat se situe dans le domaine de la vision par ordinateur et de l’apprentissage automatique. Il est lié à l'interprétation d'images et de vidéos: reconnaissance d'images, classification d'images et détection d'objets dans des images. L’un des plus grands défis de la Vision par Ordinateur est le développement de modèles permettant l'interprétation du contenu des images (objets, relations entre objets, actions, descriptions d’images en langage naturel, etc.). Le porteur du projet a développé des modèles statistiques qui ont contribué à l'essor de l'Apprentissage Statistique pour la Vision par Ordinateur.
Il est auteur ou coauteur de 27 articles dans des revues internationales et de plus de 140 articles dans de grandes conférences internationales. Selon Google Scholar, ces articles ont reçu plus de 10 000 citations pour un indice H de 45 (35 depuis 2014).
Le programme de travail de la chaire s'articulera autour de 4 axes, qui s'appliqueront tous au domaine de la Vision par Ordinateur : i) apprentissage fédéré respectueux de la vie privée ii) apprentissage profond explicable iii) algorithmes d'apprentissage en profond frugaux et efficaces iv) algorithmes d'apprentissage profonds robustes et fiables . Pour chaque direction, les objectifs, la méthodologie, l'originalité et la faisabilité du travail sont indiqués dans la proposition. Des sujets détaillés pour 5 doctorants et 4 post-doctorants sont également présentés. En ce qui concerne la faisabilité, les ressources informatiques seront fournies par le CRIANN. Le porteur du projet s’engage à consacrer 80% de son temps de recherche au projet.
En ce qui concerne les aspects financiers, le financement demandé est de 510 k€ (incluant avec les frais de gestion), soit 36% du coût total (1417 k€). La chaire sera cofinancée par Safran (33%), la Région Normandie (14%) et l'Université de Caen Normandie (17%). Le budget servira principalement à financer 5 étudiants en doctorat et 4 postdoctorants.
Coordination du projet
Frédéric Jurie (Groupe de recherche en Informatique, Image, Automatique et Instrumentation de Caen Unité de recherche)
L'auteur de ce résumé est le coordinateur du projet, qui est responsable du contenu de ce résumé. L'ANR décline par conséquent toute responsabilité quant à son contenu.
Partenariat
GREYC Groupe de recherche en Informatique, Image, Automatique et Instrumentation de Caen Unité de recherche
Aide de l'ANR 509 760 euros
Début et durée du projet scientifique :
août 2020
- 48 Mois
