2018 LIA - CNRS Workshop on Artificial Intelligence and its Applications
The workshop will be one full day.
The objective of this workshop within the LIA laboratory on "Information, Learning and Control" of the CNRS is to stimulate and facilitate discussion, interaction, and comparison of approaches, methods, and ideas related to specific topics, both theoretical and applied, in the general area of Artificial Intelligence.
The program should primarily include presentations by researchers from Ecole de Technologie Superieure (Québec, Canada) and L2S (CNRS-CentraleSupélec-Université Paris Sud) and CVN (CentraleSupélec). This workshop will provide an informal setting where participants will have the opportunity to discuss specific technical topics in an atmosphere that fosters the active exchange of ideas.
Workshop attendees do not need to register.
- 9h15 - 9h45 : Welcome & coffee
- 9h45 - 10h30
Title: Constrained-CNN losses for Weakly Supervised Segmentation
Speaker: Prof. Ismail Ben Ayed, Research Chair on Artificial Intelligence in Medical Imaging, Ecole de Technologie Superieure (ETS)
Abstract: Recently, weakly supervised learning has drawn tremendous research interests in computer vision. The purpose is to mitigate the lack of full and laborious annotations in dense prediction tasks such as semantic segmentation. In this talk, I will discuss some recent developments in this direction, focusing on how to enforce various types of constraints on convolutional neural network (CNN) outputs, which can leverage unlabeled data, guiding training with domain-speciﬁc knowledge. I will discuss several key technical aspects in the context of CNNs with partial labels, including constrained optimization, dense conditional random fields and computational tractability. I will use various illustrations and applications, and show how constrained-CNN losses can reach performances that are close to full supervision while using fractions of the full ground-truth labels.
* Kervadec et al., Constrained-CNN losses for weakly supervised segmentation, MIDL 2018
* Tang et al., On regularized losses for weakly supervised CNN segmentation, ECCV 2018
Biography: Ismail Ben Ayed received the PhD degree (with the highest honor) in computer vision from the INRS-EMT, Montreal, QC, in 2007. He is currently Associate Professor at ETS Montreal, where he holds a research chair on artificial intelligence in medical imaging. Before joining the ETS, he worked for 8 years as a scientist at GE Healthcare, London, ON, where he conducted research in medical image analysis. He also holds an adjunct professor appointment at Western University (since 2012). Ismail's interests are in optimization, computer vision, machine learning and their potential applications in medical image analysis. He co-authored a book and over seventy peer-reviewed publications, mostly published in the top venues in these subject areas. During his experience with GE, he received the GE innovation award (2010) and filed seven US patents. Dr. Ben Ayed serves regularly as program committee member for the flagship conferences of the field, and as regular reviewer for the top journals. He received the outstanding reviewer award for CVPR in 2015.
- 10h30 - 11h15
Title: Solid Harmonic Wavelet Scattering for Predictions of Molecule Properties
Speaker: Dr. Georgios Exarchakis, Département d'informatique, ENS Ulm
Abstract: We present a machine learning algorithm for the prediction of molecule properties inspired by ideas from density functional theory. Using Gaussian-type orbital functions, we create surrogate electronic densities of the molecule from which we compute invariant "solid harmonic scattering coefficients" that account for different types of interactions at different scales. Multi-linear regressions of various physical properties of molecules are computed from these invariant coefficients. Numerical experiments show that these regressions have near state of the art performance, even with relatively few training examples. Predictions over small sets of scattering coefficients can reach a DFT precision while being interpretable.
- 11h15 - 12h
Title: Spectral Matching & Learning of Surface Data - Examples on Brain Surfaces
Speaker: Prof. Hervé Lombaert, Département de génie logiciel et des technologies de l'information, Ecole de Technologie Superieure (ETS)
Abstract: How to analyze complex shapes, such as of the highly folded surface of the brain? This talk will show how spectral shape analysis can benefit general problems where data fundamentally lives on surfaces. Here, we exploit spectral coordinates derived from the Laplacian eigenfunctions of shapes. Spectral coordinates have the advantage over Euclidean coordinates, to be geometry aware and to parameterize surfaces explicitly. This change of paradigm, from Euclidean to spectral representations, enables a classifier to be applied *directly* on surface data, via spectral coordinates. Brain matching and learning of surface data will be shown as examples.
Bio: Hervé Lombaert is a Faculty at ETS Montreal - His research interests are in Statistics on Shapes, Data & Medical Images. He had the chance to work in multiple centers, including Microsoft Research (Cambridge, UK), Siemens Corporate Research (Princeton, NJ), Inria Sophia-Antipolis (France), McGill University (Canada) - more at [ http://cim.mcgill.ca/~lombaert ]
- 12h - 14h : Lunch break
- 14h -14h45
Titre: The Role of Information Complexity and Randomization in Representation Learning
Speaker: Prof. Pablo Piantanida, Laboratoire des Signaux et Systèmes (L2S), CentraleSupélec-CNRS-Université Paris Sud
Abstract: A grand challenge in representation learning is to learn the different explanatory factors of variation behind the high dimensional data. Encoder models are often determined to optimize performance on training data when the real objective is to generalize well to unseen data. In this talk, we present a sample-dependent bound on the generalization gap of the cross-entropy loss that scales with the information complexity (IC) of the representations, meaning the mutual information between inputs and their representations. The IC is empirically investigated for standard multi-layer neural networks with SGD on MNIST and CIFAR-10 datasets; the behaviour of the gap and the IC appear to be in direct correlation, suggesting that SGD selects encoders to implicitly minimize the IC. We specialize the IC to study the role of Dropout on the generalization capacity of deep encoders which is shown to be directly related to the encoder capacity, being a measure of the distinguishability among samples from their representations. Our results support some recent regularization methods.
- 14h45 - 15h30
Titre: AtlasNet: Multi-Atlas non-Linear Deep Networks for Medical Image Segmentation
Speaker: Dr. Maria Vakalopoulou, Centre de Vision Numérqiue (CVN), CentraleSupélec
Abstract: Deep learning methods have gained increasing attention in addressing segmentation problems for medical images analysis despite challenges inherited from the medical domain, such as limited data availability, lack of consistent textural or salient patterns, and high dimensionality of the data. In this paper, we introduce a novel multi-network architecture that exploits domain knowledge to address those challenges. The proposed architecture consists of multiple deep neural networks that are trained after co-aligning multiple anatomies through multi-metric deformable registration. This multi-network architecture can be trained with fewer examples and leads to better performance, robustness and generalization through consensus. Comparable to human accuracy, highly promising results on the challenging task of interstitial lung disease segmentation demonstrate the potential of our approach.
- 15h30 - 16h : Coffee break
- 16h - 16h45
Titre: Minimal Exploration in Structured Stochastic Bandits
Speaker: Prof. Richard Combes, Laboratoire des Signaux et Systèmes (L2S), CentraleSupélec-CNRS-Université Paris Sud
Abstract: This paper introduces and addresses a wide class of stochastic bandit problems where the function mapping the arm to the corresponding reward exhibits some known structural properties. Most existing structures (e.g. linear, Lipschitz, unimodal, combinatorial, dueling, ...) are covered by our framework. We derive an asymptotic instance-specific regret lower bound for these problems, and develop OSSB, an algorithm whose regret matches this fundamental limit. OSSB is not based on the classical principle of "optimism in the face of uncertainty" or on Thompson sampling, and rather aims at matching the minimal exploration rates of sub-optimal arms as characterized in the derivation of the regret lower bound. We illustrate the efficiency of OSSB using numerical experiments in the case of the linear bandit problem and show that OSSB outperforms existing algorithms, including Thompson sampling.
- 16h45 - 17h30
Titre: A Graph-based Framework for Text Categorization
Speaker: Prof. Fragkiskos D. Malliaros, Centre de Vision Numérqiue (CVN), CentraleSupélec
Abstract: With the rapid growth of social media and networking platforms, the available textual resources have been increased. Text categorization refers to the machine learning task of assigning a document to a set of two or more predefined categories (or classes). In this talk, I will present a graph-based framework for text categorization. Contrary to the traditional Bag-of-Words approach, we consider the Graph-of-Words (GoW) model in which each document is represented by a graph that encodes relationships between the different terms. Based on this formulation, the importance of a term is determined by weighting the corresponding node in the document, collection and label graphs, using node centrality criteria. We also introduce novel graph-based weighting schemes by enriching graphs with word-embedding similarities, in order to reward or penalize semantic relationships. Our methods produce more discriminative feature weights for text categorization, outperforming existing frequency-based criteria.