20-22 NOVEMBER 2019
PROGRAM
SOPHI.A SUMMIT 2019 AI WEEK
Wed. November 20th
AI & HEALTH AI & ETHICS
Wed. November 20th 9:30AM 10AM
Welcome Breakfast
OPENING CEREMONY Marco Gori, Head of Siena Artificial Intelligence Lab The Challenges of the Transition from Ai to aI
12:30PM
Lunch
AI & HEALTH 2PM
Plenary Session: Emmanuel Bacry Health Data Hub
Plenary Session: Michael Tangermann
2:40PM
Freiburg University, Dept. of Computer Science Auditory BCI as a novel treatment for stroke-induced aphasia
3:20PM
Break
Parallel Session
Parallel Session
Chair H. Delingette
Chair O. Humbert
Grand Chelem Room
Coaches Room
3:30PM
Mining Natural Language Arguments on Clinical Trials S.Villata, T. Mayer & E. Cabrio , I3S
LightHouse: An AI-Based Solution for the Blind & Visually Impaired O. Bara, Scalian
3:50PM
AI & Biophysics for Computational Cardiology: Learning by Heart M. Sermesant, Inria
PredictMed: Using AI for Medical Prediction in Neurology C.M. Bertoncelli, Lenval Hospital
4:10PM
Box-constrained Minimax Classifier for Personalized Medicine L. Fillatre, I3S
AI in Abdominal Aortic Aneurysm J. Raffort, CHU Nice , M. Carrier,, Univ. Paris-Saclay C. Adam, Univ. Paris-Saclay F. Lareyre, CHU Nice
4:30PM
A Statistical Classification Approach to Differentiate Bipolar Disorder & Major Depressive Disorder S. Barbosa, IPMC
AI to Design Odorants Targeting Emotion & Motivation D. Renaud, CHU Nice
5PM
Break
AI ETHICS ROUND TABLE 5:15PM – 6:45PM
Keynote Speaker: Lyse Langlois, Laval University & International Observatory on Societal Impacts of AI & the Digital Technology L. Godefroy, M. Teller - Université Côte d'Azur, G. Lewkowicz - Université Libre de Bruxelles, S. Cheran - HPE, K. Benyekhlef - Université de Montréal
7PM – 9PM
SOCIAL EVENT (Beachcomber Grand Chelem Room)
Thu. November 21st AI FUNDAMENTALS AI, SMART TERRITORIES & INFRASTRUCTURES IMPACT OF AI ON ENTERPRISES
Thu. November 21st AI FUNDAMENTALS Plenary Session: Frank Nielsen
9AM
Teacher-Researcher, Polytechnique Machine learning using the framework of information geometry
9:45AM
Break
Parallel Session
Parallel Session
Chair C. Bouveyron
Chair F. Precioso
Grand Chelem Room
Coaches Room
10AM
Monitoring Cyberbullying through Message Classification & Social Network Analysis E. Cabrio, S. Villata, M. Corazza, I3S
Topological and Informational Methods for AI P. Baudot, MedianTechnologies
10:20AM
Optimal Transport & AI: Learning with the Least Effort R. Flamary, Laboratoire Lagrange
Machine Learning Meets Security M. Bezzi, SAP
10:40AM
Theoretical Foundations for Interpretability in AI: the LIME case D. Garreau, LJAD
Reinforcement Learning for Navigation in Dynamic Environments A. Celani, ICTP
11AM
A Deep Text Based Recommender System M. Corneli, MSI Univ. Côte d’Azur
Explainable AI Based on the Feature Contribution S. Grah, Thales
11:20AM
Break
11:30AM
A Stochastic Blockmodel for Interaction Lengths in a Dynamic Network R. Rastelli, UCD Ireland
Spherical Convolutional Neural Network for Fiber Orientation Distribution Function & Micro-Structure Parameter Estimation from Diffusion MRI S. Sedlar et al., Inria
11:50AM
Expert in the Loop: Human-Machine Interaction to Build Safe AI-Based Systems in Healthcare M. Zuluaga, Eurecom
The Role of Network Topology for Distributed Machine Learning G. Neglia, Inria
12:10PM
Missing Data Meets Deep Learning P.A. Mattei, Inria
Explainable AI for Operational Decisions Greger Ottosson, IBM
12:30PM
Lunch & Exhibition
Thu. November 21st AI, SMART TERRITORIES & INFRASTRUCTURES 2PM
Plenary Session: Niloy Mitra University College London, Dept. of Computer Science Creative AI: Machine Learning meets Smart Content Generation
Plenary Session: Claude Le Pape 2:40PM
Schneider Electric,VP Technology Portfolio & Partnerships Optimization and Analytics Domain Leader “Wise” Use of Machine Learning for Energy and Asset Management
3:20PM
Break
Parallel Session
Parallel Session
Chair P. Alliez Grand Chelem Room
Chair N. Maïzi Coaches Room
Mobile AI: Challenges & Opportunities M. Debbah, Huawei
Towards Energy Behavior Change in Smart Buildings Using AI R. Akkouche, S. Selosse, G. Guerassimoff, Mines ParisTech
3:50PM
AI for Embedded, AI Challenges & Solutions F. Piry, ARM
Machine Learning Methods to Assist Multi-energy Systems Optimization in a Smart Grid D. Bousnina, G. Guerassimoff, Mines ParisTech
4:10PM
Using Scenarios to Explore AI Implications for Infrastructures & Territories J. Hoffmann, SKEMA
How AI can Improve the Traveler Experience R.A. Agost, Amadeus
4:30PM
Using AI to Optimize Industrial & IT Prescriptive Maintenance Operations P. Baudelle, HPE
Automatic & Robust Chain for Urban Reconstruction from Satellite Imagery Y. Tarabalka, LuxCarta
4:50PM
Making AI Happen: How to Successfully Implement AI Transformations C. Bataller, Accenture
From Satellite Images to Digitized Representation O. Tasar, Inria
3:30PM
5:15PM
Break
IMPACT OF AI ON ENTERPRISES 5:30PM 7PM
ROUND TABLE E.Viale - Accenture, G. Floyrac - Amadeus, F. Farzaneh - IPAG, P. Masse - HPE, B. Sinclair-Desgagne - SKEMA
Fri. November 22nd
AI & HEALTH
Fri. November 22nd AI & HEALTH Plenary Session: Benoît Lamarche 8:30AM
Laval University, Quebec, Faculty of Agriculture & Food Sciences Scientific Director of PULSAR Digital technology for sustainable health at Laval University
9:15AM
Break
Sessions Chair T. Papadopoulo Grand Chelem Room
9:30AM
iBiopsy® - Leveraging AI Technologies in Imaging to Unlock the Power of Precision Medicine N. Boujeema, Median Technologies
9:50AM
Source-Target Proximity Analysis in Protein Networks Using Random Walks with Restart; Applications to Cancer Drug Resistance Prediction from Single-Cell Data F. Cazals, A. Jean-Marie, D. Mazauric, G. Santa Cruz, Inria J. Roux, IRCAN
10:30AM
Clinical Setting Specific Semi-Supervised Approach to Model Synthetic Genomes of Pregnant Women to Determine Confidence Intervals for NIPT S. Bottini, MDLab, Université Côte d’Azur
10:50AM
Break
11:00AM – 1PM
CLOSING SESSION 3IA Launch
SOPHI.A SUMMIT 2019
FREE-ENTRY SIDE EVENTS
SOPHI.A SUMMIT 2019 FREE-ENTRY SIDE EVENTS AI MASTER CLASSES Tue. November 19th
8:30AM – 5PM Campus Polytech Nice Sophia 930 Route des Colles, 06410 Biot Sophia Antipolis
AI EDUCATION FORUM Tue. November 19th 10AM – 4PM
Campus SophiaTech, Espace Entreprises 930 Route des Colles, 06410 Biot Sophia Antipolis
PUBLIC CONFERENCE OF LUC JULIA Author of « Artificial Intelligence Doesn’t Exist » & Co-inventor of Siri
Sat. November 23rd 4PM – 6:30PM Palais des Congrès Antibes - Juan les Pins 60 Chemin des Sables, 06160 Antibes
SOPHI.A SUMMIT 2019 ABSTRACTS
SOPHI.A SUMMIT 2019 ABSTRACTS Opening session Wed. November 20th morning
Marco Gori The Challenges of the Transition from Ai to aI The structure of the scientific revolution was very well addressed by Thomas Kuhn (1962) in his epistemological notion of “paradigm-shift.” Are we in front of a truly paradigm-shift in artificial intelligence? What is the role of the explosive mixture of big data and nowadays huge computational resources? In this talk, I claim that, despite of the unquestionable spectacular results achieved especially in some fields, most of nowadays computational models of machine learning are quite primitive and that, like in computer vision, we are facing artificial problems harder than those offered by nature. Basically, so far, most of the progress is based on “Artificial intelligence” whereas we need “artificial Intelligence”, so as to improve the quality of cognitive processes, while cutting computational resources. This will likely spur strong innovation and might change the framework of competition amongst hi-tech companies, while the saving of energy consumption nicely fits also the need to reduce the environmental impact. Finally, the transition from Ai to aI seems to intersect the distinctive philosophy of European Union, which is strongly promoting prosperity through human-centric AI.
SOPHI.A SUMMIT 2019 ABSTRACTS
SOPHI.A SUMMIT 2019 ABSTRACTS AI & Health Wed. November 20th afternoon
Emmanuel Bacry Analysis of health data is essential to advance research and enlighten the decision-maker or the citizen. For many reasons, these are now underutilized. The obstacles are not technical, but rather organizational and cultural. Overcoming them requires to acknowledge our collective responsibility on a founding principle: health data financed by french national solidarity constitute a national asset. These data must be put at the service of the common good in respect of the ethics and fundamental rights of citizens. It is essential to guarantee an easy, unified, transparent and secure access. This is the role of the Health Data Hub.
Michael Tangermann Information on how well a stroke patient performs an exercise is easy to obtain for most behavioural tasks, as motor performance can be measured. For cognitive tasks such a direct metric is harder to obtain, which makes it difficult for a therapist to support the (re-)learning of a good cognitive strategy. If open behaviour is not accessible, a brain-computer interface (BCI) system can be used to analyse ongoing brain signals. Utilizing machine learning methods, task-related signal features can be analysed and informative feedback can be provided to the patient. In a collaboration between my lab and partners from neurology, we have put this idea to work with several chronic stroke patients in a novel cognitive rehabilitation paradigm. It makes use of an auditory BCI setup to support the rehabilitation from expressive aphasia. The clinical effect obtained with this BCI-supported training is high compared to standard language therapy and it generalizes beyond the trained task. Based on training-induced changes in the electrophysiology and resting state connectivity, we are confident, that our rehabilitation approach is capable to reinforce individual language-related processes in these stroke patients.
SOPHI.A SUMMIT 2019 ABSTRACTS AI & Health Wed. November 20th afternoon Parallel Session Serena Villata, Tobias Mayer, Elena Cabrio (I3S, CNRS / UniversitÊ Côte d’Azur). Evidence-based decision making in the health-care domain targets at supporting clinicians in their deliberation process to establish the best course of action for the case under evaluation. Although the reasoning stage of this kind of frameworks received considerable attention, little effort has been devoted to the mining stage. To fill this gap, we propose a deep bidirectional transformer approach combined with different neural networks to address the AM tasks of component detection and relation prediction in Randomized Controlled Trials, and we evaluate this approach on a new huge corpus of 500 abstracts from the MEDLINE database. Maxime Sermesant: Electromechanical models of the heart have made important progress over the last decades, and personalized models can now be envisaged for clinical applications. However there are still challenges in translating such models to the clinics. On the other hand, the recent progress in computing power and available data makes it possible to develop accurate data-driven approaches for healthcare but such artificial intelligence approaches often lack of robustness. Machine learning and biophysical modelling are very complementary approaches, with biophysical models offering a principled way to introduce physiological constraints. In this talk I will present results on personalized electromechanical models of the heart and research where we combined biophysics and AI in different ways in order to leverage their strengths. Different clinical applications in computational cardiology will be presented.
SOPHI.A SUMMIT 2019 ABSTRACTS AI & Health Wed. November 20th afternoon Lionel Fillatre: A clinical decision support system (CDSS) is an instrument that is designed to provide physicians with assistance with clinical decision-making tasks. The successful development of statistical classifier-based CDSS is often impaired by one of several hurdles: unbalanced training sets, change of proportions between training and testing sets, specific cost function, mixture of categorical and continuous variables. To address these issues, we have designed a box-constrained minimax classifier in which continuous variables are discretized. Experiments performed on medical datasets confirmed that our classifier outperformed other standard classifiers when considering changes in class proportion, and tended to equalize the class-conditional risks.
Suzana Barbosa: The aim of this study was to identify blood biomarkers that discriminate Major Depressive Disorder (MDD) from Bipolar disorder (BD) patients when in a depressed state. 133 patients were included, and blood was assessed for 41 biomarkers. A nonparametric bootstrap combined with a regularized logistic regression model was used to identify variables allowing for differential diagnosis. After adjustment for covariates, several cytokines including interleukin TNF and IL-10 discriminated MDD from BD patients. Should the diagnostic value of these biomarkers be validated in another cohort, it will accelerate the development of a blood-based diagnostic tool to support clinical management of patients.
SOPHI.A SUMMIT 2019 ABSTRACTS AI & Health Wed. November 20th afternoon Parallel Session II Ouassim Bara: "LightHouse" is an application designed to offer a unique experience to the visually impaired or blind person and contribute to their well-being. It stands on the latest advances in Artificial Intelligence (AI), especially in one of its subfields, namely, deep learning. The application allows anyone in possession of a smartphone and without the use of one’s eyesight to know approximately what's in front of them, what object to look for and provide instructions on how to find it. The innovation of the algorithm stands not only on the latest object detection and localization algorithm, YOLO, but more importantly, on the many features (modes) that would assist the visually impaired people in their daily life and inside their home. Carlo Mario Bertoncelli: The domain of our study is AIbased diagnostics and Health analytics. We designed a supervised learning model named ‘PredictMed’. PredictMed uses a Logistic Regression algorithm for binary and multinomial classification and it is better suited to solve classification problems where the class of each patient (or group) is known a priori. First in literature, PredictMed has been developed and validated to predict scoliosis development and to identify factors associated with autistic features, intellectual disabilities, feeding disorders in cerebral palsy series. The predictive performance in terms of accuracy, sensitivity, specifity scored between 74% and 95%. This presentation discusses the potential of using PredictMed, for health-related outcomes prediction and presents a few examples.
SOPHI.A SUMMIT 2019 ABSTRACTS AI & Health Wed. November 20th afternoon Juliette Raffort, Marion Carrier, CÊdric Adam, Fabien Lareyre: Abdominal aortic aneurysm (AAA), which corresponds to a focal dilatation of the aorta, is associated with extremely high rates of mortality. Artificial intelligence (AI) has brought new insights in the management of cardiovascular diseases but its interest in AAA has been so far poorly described. The aim of this report was to summarize current knowledge and potential applications of AI in AAA. A comprehensive literature review was performed and identified three main topics including imaging analysis, characterization of AAA morphology and fluid dynamics, development of predictive programs. Based on the published studies, current limits and future directions are discussed. David Renaud: The use of olfactory agents has shown promising results in reducing levels of anxiety or depressive symptoms as well as stimulating individuals with high levels of amotivation. Depression, anxiety and amotivation are core behavioural symptoms in Alzheimer’s disease and related disorders. Our project aims to connect machine learning algorithms and molecular modeling with properties measured through sensory analysis and psychophysiology experiments on human individuals. The goal is to design odorants to fight depression and anxiety using non-pharmacological approaches.
SOPHI.A SUMMIT 2019 ABSTRACTS AI & Health Wed. November 20th afternoon Round Table “AI and ethics” Lyse Langlois, Lémy Godefroy, Marina Teller, Gregory Lewkowicz, Sorin Cheran, Karim Benyekhlef One of the major challenges in the development of the uses of Artificial Intelligence is the future of individuals and societies. This round table will address the ethical and legal issues raised by this new technology. Giving as much weight to ethics as to law as mechanisms of social regulation makes it possible to better feed reflection, debate collectively and anticipate a possible democratic deficit induced by these developments.
SOPHI.A SUMMIT 2019 ABSTRACTS
AI Fundamentals Thu. November 21st morning
Frank Nielsen Energy, loss functions, and distances lie at the heart of the design and evaluation of learning machines. We review the underlying dualistic information geometry induced by principled classes of distances, explain the criteria and properties of statistical invariance, and show how information projections and natural gradient explain key mechanisms of geometric (deep) learning. References: - An elementary introduction to information geometry, arXiv:1808.08271 (2018) - Clustering in Hilbert simplex geometry, arXiv:1704.00454 (2017) - Relative Fisher information and natural gradient for learning large modular models, ICML (2017)
SOPHI.A SUMMIT 2019 ABSTRACTS AI Fundamentals Thu. November 21st morning Parallel Session I Elena Cabrio, Serena Villata, Michele Corazza: In online social media, each episode of online activity aimed at offending, menacing, harassing or stalking another person can be classified as a cyberbullying phenomenon. Given its societal impact, the implementation of cyberbullying detection systems, combining abusive language detection and social network analysis, has attracted a lot of attention in the last years. However, the adoption of such systems in real life is not straightforward, and given the negative effects misleading analyses could have on potential abusers and victims, a transparent approach should be adopted, in which cyberbullying identification should be mediated by human judgment. In this talk, we report on our experience on CREEP, an EIT funded project on cyberbullying effects prevention. In particular, we will present the project outcome, i.e. a platform for the monitoring of cyberbullying phenomena on social media. The system aims at supporting supervising persons at identifying potential cases of cyberbullying through an intuitive, easy-to-use interface. We evaluate the hate speech detection classifier on a set of manually annotated data from several social media platforms, and for three languages.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Fundamentals Thu. November 21st morning RĂŠmi Flamary: Thanks to recent optimization procedures, Optimal Transport (OT) has been recently reintroduced with success to the machine learning and Artificial Intelligence community. This presentation will be a short introduction to the fundamental mathematical problem of optimal transport followed by some illustrations of applications in machine learning and Artificial Intelligence. The first part of the presentation will introduce the theory of optimal transport and the optimization problem that will be illustrated on simple examples. The second part of the presentation will discuss recent applications of OT in the field of machine learning and artificial intelligence. We will see how optimal transport can be used to perform domain adaptation, data processing for data seen as histograms and graphs and train generative adversarial networks. D. Garreau: Artificial Intelligence methods are used more and more often for sensitive applications, sometimes replacing humans in critical decision-making processes. These decisions have direct impact on the health and finances of the (mostly involuntary) users, and it is no longer acceptable to gauge an algorithm solely by its performances. Indeed, it is difficult to accept a decision taken by some entity that cannot explain how it came to this decision, even if it is more accurate on the long run. Interpretability has therefore become a considerable challenge for the future of the methods used in Artificial Intelligence. In this communication, we focus on a local notion of interpretability: how did the model predict a value for this specific example? A method designed to provide local explanation for such complicated models has been proposed in 2016, LIME (Local
SOPHI.A SUMMIT 2019 ABSTRACTS AI Fundamentals Thu. November 21st morning Interpretable Model-Agnostic Explanation). However, if the explanations provided by LIME seem reasonable in practice, there is no theoretical evidence supporting its use. In particular, there is no connection between the prediction of the global model and that of the surrogate. In this communication, we propose to provide such a link in a simplified setting. By doing so, we highlight the importance of the regularity of the model around the example, which allows us to give insight as to how the scale of the neighborhood should be set (a crucial parameter of the method).
Marco Corneli: An ordinal data matrix is considered, whose entries are notes that customers give to some products. The rows of the matrix correspond to the users, the columns to the products. A possibly large amount of missing data is allowed. We introduce a novel deep learning architecture to predict the missing entries in the ordinal data matrix. This architecture benefits of the textual information contained in the reviews that users may formulate regarding the products they rank. Based on a variational autoencoder and a non-linear Latent Dirichlet allocation, our framework also allows to forecast the most likely words that would be used by an user to review a product. Some experiments on simulated are real datasets are reported to highlight the features of the proposed methodology.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Fundamentals Thu. November 21st morning Riccardo Rastelli: In this talk I will introduce a new stochastic blockmodel that focuses on the analysis of interaction lengths in dynamic networks. The model does not rely on a discretization of the time dimension and may be used to analyze networks that evolve continuously over time. The framework relies on a clustering structure on the nodes, whereby two nodes belonging to the same latent group tend to create interactions and noninteractions of similar lengths. Inference is performed using a fast variational expectation-maximization algorithm, and a widely used clustering criterion is adopted to perform model choice. I will conclude the talk with a demonstration of the methodology on artificial data and on a dataset concerning face-to-face interactions between high-school students. Maria Zuluaga: In this talk, I will present my research on the development of reliable learning frameworks addressing the problems of data complexity and low error tolerance in high-risk applications. I will discuss Interactive Machine Learning as a way to optimize the learning process of AI systems in the presence of complex data. Interactive ML includes aspects of multiple learning techniques (e.g. Reinforcement and Semi-supervised learning, Transfer, Online and Active learning), under the idea of integrating an expert agent in the loop. Achieving efficient interactivity passes by continuously measuring a system’s performance and informing the expert about potential anomalous behaviors. I will present on-going work on performance assessment in the absence of ground truth, and on how to exploit these measurements to provide evidence for informed decision-making to the expert.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Fundamentals Thu. November 21st morning Pierre-Alexandre Mattei: We consider the problem of handling missing data with deep latent variable models (DLVMs). First, we present a simple technique to train DLVMs when the training set contains missing-at-random data. Our approach, called MIWAE, is based on the importance-weighted autoencoder (IWAE), and maximises a potentially tight lower bound of the log-likelihood of the observed data. Compared to the original IWAE, our algorithm does not induce any additional computational overhead due to the missing data. We also develop Monte Carlo techniques for single and multiple imputation using a DLVM trained on an incomplete data set. We illustrate our approach by training a convolutional DLVM on incomplete static binarisations of MNIST. Moreover, on various continuous data sets, we show that MIWAE provides extremely accurate single imputations, and is highly competitive with stateof-the-art methods. This talk is based on the paper available at: http://proceedings.mlr.press/v97/mattei19a.html
SOPHI.A SUMMIT 2019 ABSTRACTS AI Fundamentals Thu. November 21st morning Parallel Session II Pierre Baudot: As a domain that formalizes the classification and recognition of patterns-structures in mathematics, Topological Data Analysis has progressively gathered the interest of the data science community. On the side of neural networks following Hinton, Amari, information geometric approaches have provided well defined metric and gradient descent methods. This presentation will focus on an original approach of algebraic topology intrinsically based on probability/statistics and information, developed notably with D. Bennequin since 2006. Information topology characterizes uniquely usual information functions, unraveling that two theories, cohomology and information theory, are of the same nature. These probabilistic tools describe the statistical forms or patterns present in databases and make them correspond to discrete symmetries. The set of statistical interactions-dependencies between k elementary variables is quantified by the multivariate mutual information between these k components. It provides a generalized and metric-free decomposition of free energy that is used in machine learning and artificial intelligence, with brand new computationally expensive algorithm for unsupervised and supervised learning, where simplicial complexes provide topologically constrained Deep Neural Networks architectures. Its application to gene expression under open source software makes it possible to detect functional modules of covariant variables (collective dynamics) as well as clusters (corresponding to condensation phenomena and negative synergistic interactions) in high dimension, and thus to analyze the structure and to quantify diversity in data or arbitrary complex systems.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Fundamentals Thu. November 21st morning Michele Bezzi: The intelligent enterprise needs intelligent security. Learn how SAP security research applies machinelearning techniques to improve security by automatically discovering security information in the dark web, classifying security-relevant code in software repositories, or detecting new threats. You will also understand how machine learning algorithms can expose to security and privacy risk, and how we can make machine learning more secure. Antonio Celani: In reinforcement learning, an animal maximizes its long-term reward by taking actions in response to its external environment and internal state. Learning occurs by reinforcing behaviour based on feedback from past experiences. Similar ideas have been used to develop intelligent agents that have achieved spectacular performance in strategic games such, visual-based video game play and robotics. In the field, however, constraints imposed by variable and uncontrolled conditions prevent learning agents from using data intensive learning algorithms and the optimization of model design needed for quicker learning. These are the conditions most often faced by living organisms. A striking example in nature is provided by thermal soaring. Atmospheric convection is not consistent across days and, even under suitable conditions, the locations, sizes, durations and strengths of nearby thermals are unpredictable. Gliders and birds operate at spatial and temporal scales where fluctuations in wind velocities last a few seconds and may mask or falsely enhance a glider’s estimate of its mean climb rate. How soaring birds find and navigate thermals within this complex landscape is unknown. Reinforcement learning provides an appropriate framework in which to identify an effective navigational strategy as a sequence of decisions made in response to environmental cues. Here we show how to use reinforcement learning to train a glider in the field to navigate atmospheric thermals autonomously.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Fundamentals Thu. November 21st morning Simon Grah: Symbolic AI and Machine Learning are booming, and have many business applications. However, they present a problem: their reasons are difficult, if not impossible, to understand. This need is important. Indeed, for the engineer who builds the AI, it allows to understand how it works, and therefore to debug it if necessary. For an end user, this will allow him to better understand the AI, and thus increase his confidence in it and provide more relevant decision support. To meet this need, a new branch, the explainable AI, has been developing in recent years. Explanations can be global, i.e. we can try to explain the general behavior of the AI, or local, when we try to explain a prediction of the AI. In this presentation, we focus on the latest types of explanations. We will use Shapley's values, which come from Collaborative Game Theory (CGT). But there are two difficulties in the use of the Shapley value in AI. Firstly, its computation complexity increases exponentially with the number of features. We will present here different ways to approximate them, especially in a Machine Learning context. Secondly, one needs to assess the worth allotted to each subset of features, that is to assess the predicted value taking into account only some features. These assessments require some assumptions. For instance, some techniques assume the independence between the features, some others asks Gaussian features, etc. In this Machine Learning context, we will also be interested in their calculation when we no longer consider a feature-by-feature calculation, but a group of features per group of features. We will show the performance in term of error and in term of computation time of each approximation on some simulated data. Then, we will present the results on Thales use cases. These use cases may involve medical or industrial data.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Fundamentals Thu. November 21st morning Sara Sedlar, Abib Alimi, ThĂŠodore Papadopoulo, Rachid Deriche, Samuel Deslauriers-Gauthier: Diffusion Magnetic Resonance Imaging (dMRI) is an imaging technique that can reveal underlying tissue properties by capturing diffusion of water molecules. In the case of dMRI scans of the central nervous system, two important ways of describing the tissue are through the fibre orientation distribution function (fODF), also known as FOD, and micro-structure parameter estimation. Convolutional neural networks are proven to be a powerful tool for many computer vision problems where the data is acquired on a regular grid in Euclidean space. As the dMRI signals used in our experiments are acquired on spheres, we have investigated how the spherical CNN model can be adapted to properties of dMRI data. Giovanni Neglia: Many learning problems are formulated as minimization of some loss function on a training set of examples. Distributed gradient methods on a cluster are often used for this purpose. In this paper, we study how the variability of task execution times at cluster nodes affects the system throughput. In particular, a simple but accurate model allows us to quantify how the time to solve the minimization problem depends on the network of information exchanges among the nodes. Interestingly, we show that, even when communication overhead may be neglected, the clique is not necessarily the most effective topology, as commonly assumed in previous works.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Fundamentals Thu. November 21st morning
Greger Ottosson: As we apply Machine Learning to increasingly critical use cases, explainability as an afterthought is no longer sufficient. Already during model training we need to consider how to deliver on transparency. Should we restrict our options to Machine Learning methods that are intrinsically transparent, like simple regression, decision trees or rule lists? Or leverage any opaque ensemble- or deep learning technique that performs the best, and aim to fix explainability post-hoc with LIME, Shapley or surrogate models? In this talk, we will explore what to consider for operational decisions - think finance, insurance and health care - where predictive ML-based scores are often blended with policy and strategy rules. In these systems, we need to consider prediction and explanation as a deliberately chosen pair.
SOPHI.A SUMMIT 2019 ABSTRACTS
SOPHI.A SUMMIT 2019 ABSTRACTS
AI Smart Territories & Infrastructures Thu. November 21st afternoon
Niloy Mitra The ability to generate novel, diverse, and realistic 3D shapes along with associated part semantics, movement, and structure is central to many applications requiring high-quality dynamic 3D assets or large volumes of realistic training data. A key challenge towards this goal is how to accommodate diverse shape, including both continuous deformations of parts as well as structural or discrete alterations which add to, remove from, or modify the shape constituents and compositional structure. While volumes of high-quality 3D models are available, they are often represented in a variety of different representations (images, point sets, meshes, rigging structures) and hence traditional machine learning tools cannot be directly applied to handle them. In this talk, I will discuss our recent results on a unified learning structure both for analysis and generative tasks. We will show application-specific use cases in the context of 3D model synthesis, appearance control, physical simulation, and garment animation. For more details, please visit http://geometry.cs.ucl.ac.uk/index.php.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Smart Territories & Infrastructures Thu. November 21st afternoon
Claude Le Pape In many domains (energy and water plants and networks, building management, transportation, etc.), the deployment of integrated optimization solutions requires the implementation of analytical applications, based on modelling, operations research, and artificial intelligence technologies. Relevant usages of these technologies include performance evaluation, production/consumption forecasting, anticipation of maintenance requirements, optimization of investments, operational planning and control, etc. In particular: The Internet of Things brings together multiple data on the use, the environment of use and the control of the assets composing complex systems, data which can be transformed into useful information on the state of these assets. This enables multiple actors to (i) anticipate the ageing of assets, (ii) detect the first signs of an anomaly and (iii) modify the maintenance plan accordingly. Data on energy consumption and production can be used to plan and control the use of complex systems. Machine learning techniques are extremely useful in this area, whether it is to forecast consumption, production or prices, manage uncertainties, or "understand" the varying preferences of users. To reap the benefits and avoid mistakes, solution designers must however use machine-learned models wisely. Simple examples will be given to illustrate potential errors and expose means to avoid them.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Smart Territories & Infrastructures Thu. November 21st afternoon Parallel Session I Merouane Debbah: Fueled by the availability of more data and computing power, recent breakthroughs in cloud-based machine learning (ML) have transformed every aspect of our lives from face recognition and medical diagnosis to natural language processing. However, classical ML exerts severe demands in terms of energy, memory and computing resources, limiting their adoption for resource constrained edge devices. The new breed of intelligent devices requires a novel paradigm change calling for distributed, low-latency and reliable ML at the wireless network edge. This talk will explore the potential of the Mobile AI paradigm to unlock the full potential of 5G and beyond. FrĂŠdĂŠric Piry: AI has many implications that at a first glimpse look incompatible with constraints of the Embedded world, let alone wearables: - AI requires significant computing resources, inducing significant requirements on power / energy. Offloading is often looked at the solution but this creates other challenges, as per the following points. - AI requires significant amount of data. Constant progresses are being made on that field, with inference network using data format as small as 8 or even 4b. Yet, offloading requires transferring this data to a companion device or to the cloud, increasing the amount of energy globally required.. - AI applications often require low latency, limiting task offload in some cases. The goal of our study is to apprehend the compromises between precision and data set, discuss SW partitioning between end device, companion device and edge compute, and all energy and latency implications induced by these choices.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Smart Territories & Infrastructures Thu. November 21st afternoon Jonas Hoffmann: Many organizations and territories are strengthening their abilities to cope with turbulent, unpredictable, and ambiguous settings by using Scenario Planning (Ramirez et al. 2017). I will present process and outcomes of an ongoing Scenario project on the ‘future of mobility, urban infrastructure and social implications in Europe’. Discussion will center on how actors in ‘smart territories’ benefit from engaging in future oriented multi-stakeholder dialogues to co-create values, design resilient strategies, and craft novel institutional arrangements. Pierre Baudelle: Using Artificial Intelligence (AI) to optimize industrial and IT prescriptive maintenance operations. Machine learning and deep learning techniques can be leveraged to proactively save time and money in predicting equipment failure. However prediction adoption can be limited by scalability ,interpretation and integration challenges. In this session, we will discuss how a common, open framework and analytics methodology for Prescriptive maintenance from HPE can address these challenges through two differentiated use cases: industrial operations and IT operations (AIOps). The approach will be illustrated with our experiences delivering proof-of-value services and solutions from edge to cloud.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Smart Territories & Infrastructures Thu. November 21st afternoon
Cyril Bataller: AI offers many opportunities to change the way we work and live, but getting from prototype to production is hard work. That's where Applied Intelligence comes in: It's a unique approach, combining AI with data, analytics and automation under a bold strategic vision to transform businesses not in silos, but across every function and every process, at scale. It enables organizations to do things differently and do different things. It's about embedding intelligence at the core of business to drive transformative outcomes. We will review several examples of successful operational AI implementations at scale in various industries.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Smart Territories & Infrastructures Thu. November 21st afternoon Parallel Session II R. Akkouche, S. Selosse, G. Guerassimoff: Sensors that are incorporated in Smart buildings can give access to a considerable amount of data concerning different parameters such as temperatures, humidity, air quality, and presence. These data if well processed and analyzed through Artificial Intelligence can give a clear idea about people’s energy use inside buildings. The aim of the proposed work is to apply machine learning models on sensors data and weather data in order to predict energy consumption, define its patterns and lately determine adequate energy saving recommendations. The trained and tested model can also be used to predict energy consumption of other buildings with the same parameters. In future work, the tested machine learning models can be used to detect anomalies in energy use and behaviors that lead to energy waste in order to change them through behavioral economics concepts. D. Bousnina, G. Guerassimoff: This PhD research work, which is part of the Nice Meridia Smart Energy project, focuses on the coordinated optimization of the multiple energy vectors of the smart grid (electricity, heating and cooling). It aims at developing planning and control optimization strategies for the multi-energy systems of this smart grid to maximize selfconsumption, value the most the energy flexibility potentials, minimize operational costs and Greenhouse Gas emissions of the district, and alleviate the energy bills of end users. The complexity of such a problem makes it difficult to model in a traditional way. Hence, besides mathematical optimization approaches, Machine learning methods are investigated to make predictions of future energy demand and generation and to make energy optimization and flexibility qualification based on datadriven approaches.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Smart Territories & Infrastructures Thu. November 21st afternoon Rodrigo Acuna Agost: The traveler of today has evolved, as global travel grows, and the choices of destinations, products and services seems endless. Now, modern travelers can be overwhelmed by choice and want a travel experience that suits their individual needs. However, predicting what travelers will choose, and understanding their motivations, remains a challenge especially when search functionality is often limited. In this presentation we describe how AI is transforming the travel industry by impacting different steps of the passenger journey and thus improving the overall traveler experience. We will discuss the emerging applications and challenges of AI on travel with examples of different uses cases explored by our AI research team. Yuliya Tarabalka, Sebastien Tripodi, Liuyun Duan, Frederic Trastour, Veronique Poujade, Lionel Laurore, Justin Hyland: Automatic city modeling from satellite imagery is a popular yet challenging topic in remote sensing, driven by numerous applications such as telecommunications and urban management. In this talk, we present an automated chain for large-scale 3D reconstruction of urban scenes from satellite images. The proposed framework relies on two key ingredient. First, from a stereo pair of images, we estimate a digital terrain model and a digital height model, by using a novel set of feature descriptors based on multiscale morphological analysis. Second, inspired by recent works in machine learning, we extract in an automatic way contour polygons of buildings, by adopting a convolutional network U-Net followed by a polygonization of the predicted mask of buildings. We demonstrate the potential of our chain by reconstructing different areas of the world.
SOPHI.A SUMMIT 2019 ABSTRACTS AI Smart Territories & Infrastructures Thu. November 21st afternoon Onur Tasar, Pierre Alliez, Yuliya Tarabalka, SĂŠbastien Clerc, Alain Giros: One of the most popular challenges in the field of remote sensing is to generate digitized/polygonized representations from satellite images to be included into various GIS applications. One way to generate the aforementioned digitized representation is to decompose the task into two consecutive sub-tasks, where the first one consists in generating pixel-wise maps in raster format from satellite images using advanced machine learning techniques, and the second one aims at vectorizing the obtained maps by applying computational geometry methods. In the machine learning part, we study various AI related problems such as incremental learning, domain adaptation with generative adversarial networks, and multi-task learning for semantic segmentation and pan-sharpening. In the computational geometry part, we present a novel mesh approximation method that vectorizes binary classification maps for the building class.
SOPHI.A SUMMIT 2019 ABSTRACTS
AI & Health Fri. November 22nd morning
BenoĂŽt Lamarche PULSAR: the use of digital technology for sustainable health at Laval University PULSAR at UniversitĂŠ Laval is a new collaborative research and knowledge-sharing space dedicated to interdisciplinary scientific projects. Designed and developed based on the best practices in security and data management, PULSAR provides users with access to various knowledge and data sharing services and tools, as well as the use of the resulting information. By mobilizing numerous stakeholders in the Quebec City area (researchers, decision-makers, clinicians, citizens, etc.), PULSAR will make it possible to assess all the factors that have an impact on sustainable health by advocating a holistic approach based on big data and digital science. With PULSAR, Laval University and the Quebec City area aim to position themselves in the field of open research and big data in both Quebec and Canada, as well as internationally.
SOPHI.A SUMMIT 2019 ABSTRACTS AI & Health Fri. November 22nd morning Nozha Boujeema: Median technologies is developing innovative solutions at the crossroad of machine learning, computer vision and big data analytics. The combination of imaging technologies, AI and data sciences has the potential to unleash the power of precision medicine, providing insights into novel therapies, drug development, optimal treatment strategies and enhanced clinical decision. In this context, robustness of AI technology is therefore critical, and this calls for trustworthy and responsible AI services, and for the control and regulation of algorithms prior to any deployment in clinical routine. F. Cazals, J. Roux, A. Jean-Marie, D. Mazauric, G. Santa Cruz: Noise in biological systems is a major hurdle in the singlecell multi-omics era. To recover meaningful signal in large and heterogeneous omics datasets, we introduced a novel approach to source-target proximity analysis in protein-protein-interaction networks. The sources being genes/proteins of highly variable expression and the targets being the proteins that drive cell response to cancer drugs, we compute the proximity between a source and all targets using the stationary distribution of a suitable random walk on a molecular interaction network. This novel analysis framework is aimed at discovering new sources causing drug resistance, so far confounded in gene expression noise.
SOPHI.A SUMMIT 2019 ABSTRACTS AI & Health Fri. November 22nd morning Silvia Bottini: Non-Invasive Prenatal Testing (NIPT) techniques consists in the detection of fetal aneuploidies from plasma sequencing of pregnant women. However its accuracy relies on the estimation of the amount of the fetal DNA and the sequencing depth. Here we present GenomeMixer, a semisupervised approach to create synthetic genomes to assess how these two parameters affect NIPT reliability. We trained our method on a cohort of 377 samples including 10 aneuploidies and we tested its performances on a cohort of 900 samples including 35 aneuploidies. Overall, our analysis show effective guidelines to determine criteria for confident clinical setting specific NIPT.