Prof. Jenny Benois-Pineau
University of Bordeaux, 351, crs de la Libération
Title: Explainable AI in the problems of Image Classification
Abstract: The recent focus of Artificial Intelligence (AI) educators, researchers and practitioners on the supervised learning approaches, and particularly on Deep Learning, resulted in considerable increase of performance of AI systems, but also raised the question of the trustfulness and explainability of their predictions for human decision makers and adopters. Explainable AI (XAI) is addressing this challenge by developing methods to “understand” and “explain” to humans how these systems produce their decisions.
Image classification has a wide range of applications including critical ones : such as medical imaging or security. This is why explanation of classification decisions for images in critical domain is of primarily importance for the human experts which are users of such systems. The explanation of Deep Neural Network classifiers' decisions for images consists in producing the so-called explanation maps, or ’saliency « maps containing information of the importance of pic-els and regions for the decision of the network classifier.
In this tutorial we will present such kind of methods and their use in the increasing of trustfulness of AI tools for humans. At the beginning we will formulate explanation problem. Then we will present the taxonomy of explanation methods of two large families : « black-box » and « white box « ones. We will then introduce the most popular ones which today serve as base line methods for the intensively developing research in Explainable AI (XAI).
We will also show examples of results - explanation maps which are obtained by such methods. A specific focus will be on the gradient-free methods recently developed in our research. Today, a quantity of XAI methods are available in OpenSource, hence it is interesting to give some hints to the the use of these tools for young researchers and PhD students.
In the follow up of the tutorial we will consider the question of the quality of these methods and introduce metrics for their comparison. Correctness, plausibility, stability- these properties of the methods will be discussed. An important question, which is in the focus of community now is how to include human in the loop for the assessment of taxi methods.
Finally we will present a with a large coverage of contributors form European and North American research laboratories which has been recently published by Elsevier.
Biography: Jenny Benois-Pineau is a professor of exceptional class in Computer Science at the University Bordeaux and chair of Video Analysis and Indexing research group in Image and Sound Department of LABRI UMR 58000 Université Bordeaux/CNRS/IPB. She is also chair of International relations at School of Science and Technology of University of Bordeaux comprising 9.500 students. Her topics of interest include artificial intelligence in image and multimedia analysis. She is the author and co-author of more than 240 research papers in international journals, conference proceedings, books and book chapters. She has tutored and co-tutored 26 PhD students. She is area editor of Signal Processing: Image Communication, Elsevier, member of Multimedia Tools and applications editorial board, Springer, senior associated editor JEI SPIE journals. She has served in numerous program committees in international conferences and organised workshops at: ACM MM, CIVR, CBMI, EI, MMM, IEEE ICIP, IEEE/IAPR ICPR and others and gave invited lectures in the universities of USA, Europe, Israel, Mexico. She has been coordinator or leading researcher in EU – funded and French national research projects and numerous projects with French Industrial companies. She is a member of IEEE TC on IVMSP. She was decorated by Knight of Academic Palms grade.
Prof. Rostom Kachouri
Department of Computer Science, ESIEE Paris, CNRS LIGM Laboratory, Gustave Eiffel University, France.
Title: Retinal images : From biometrie to early diagnosis of ocular pathologies
Abstract: Biometric recognition is applied especially in many highly secure environments (safe, military base, etc.). Several studies have evaluated the performance of the most common biometrics such as face, fingerprints, iris, retina, …. Some morphological biometric systems still easily defeatable using articial vectors. However, the sensory layer of the eye, Retina, has a rich and unique texture even in twins. It is also an internal organ which minimizes injuries and fraud. These properties have made the retina biometrics an active research area. In this talk, I will briefly present our proposed methods for the various stages of the retinal biometric recognition, from analysis and pretreatment of the retinal image to the decision through the characterization.
In the other hand, the World Health Organization (WHO) estimates that 285 million people are visually impaired worldwide, with 39 million blinds. Thus, there is an active effort to create and develop methods to automate screening of retinal diseases such as Glaucoma, Age-related Macular Degeneration (AMD), and Diabetic Retinopathy (DR). In specific, I will share in this talk, our recent research works, concerning earlier screening and diagnosis of the most severe ocular diseases in retinal images. A special focus on Glaucoma and RD diseases is given. In this context, the talk will highlights particularly our proprosed approaches and methodology based on Deep Convolutional Neural Network (DCNN) and Generative Adeversial Networks (GAN). Finally, I will discuss the research directions towards the use of fundus image captured by some based-Smartphone camera devices, with respect to the clinical employment.
Rostom Kachouri received his Engineer and Master degrees from the National Engineering School of Sfax (ENIS) respectively in 2003 and 2004. In 2010, he received his Ph.D. in Image and Signal Processing from the University of Evry Val d’Essonne.
From 2005 to 2010, Dr. Kachouri was an Assistant Professor at the National Engineering School of Sfax (ENIS) and then at the University of Evry Val d’Essonne. From 2010 to 2012, He held a post-doctoral position as part of an industrial project with the high technology group SAGEMCOM.
Dr. Kachouri is currently Associate Professor in the Computer Science Department and Head of apprenticeship computer and application engineering at ESIEE, Paris. He is member of the Gaspard-Monge computer science Laboartory (LIGM), unité mixte de recherche CNRS-UMLPE-ESIEE, UMR 8049.
Dr. Kachouri's current research interests include AI based Intelligent System for Retinal Pathology Screening.