Seminars 2017-2018:

October 12 (Amphi 018, RDC bâtiment UFR IM2AG)*: audio-visual analysis for human-robot interaction by Radu Horaud (INRIA Grenoble)

Robots have gradually moved from factory floors to populated spaces. Therefore, there is a crucial need to endow robots with communicative skills. One of the prerequisites of human-robot communication (or more generally, interaction) is the ability of robots to perceive their environment, to detect people, to track them over time, and to identify communicative cues, such as “who looks at whom” and “who speaks to whom”. Therefore we are interested in analysing situations in which several people are present, understand their activities, estimate who speaks and who doesn’t, etc. For that purpose we combine computer vision, audio signal processing and machine learning methods. We will briefly present the research that we carried out within this topic, and stress the importance of learning from sensory data.

Related articles
- Audio-Visual Speaker Diarization Based on Spatiotemporal Bayesian Fusion
- Tracking a Varying Number of People with a Visually-Controlled Robotic Head
- Audio-Visual Fusion for Human-Robot Interaction (slides)
 

October 26 (Auditorium Bâtiment Imag)*: Toward Interactive User Data Analytics by Sihem Amer-Yahia (CNRS, LIG, Grenoble)

User data can be acquired from various domains. This data is characterized by a combination of demographics such as age and occupation and user actions such as rating a movie, reviewing a restaurant or buying groceries. User data is appealing to analysts in their role as data scientists who seek to conduct large-scale population studies, and gain insights on various population segments. It is also appealing to novice users in their role as information consumers who use the social Web for routine tasks such as finding a book club or choosing a restaurant.  User data exploration has been formulated as identifying group-level behavior such as Asian women who publish regularly in databases. Group level exploration enables new findings and addresses issues raised by the peculiarities of user data such as noise and sparsity. I will review our work on one-shot and interactive user data exploration. I will then describe the challenges of developing a visual analytics tool for finding and connecting users and groups.

Related article
- Toward Interactive User Data Analytics


November 16 (salle 320 bâtiment UFR IM2AG): Attribution modeling increases efficiency of bidding in display advertising by Eustache Diemert (Criteo Research, Paris)

Predicting click and conversion probabilities when bidding on ad exchanges is at the core of the programmatic advertising industry. Two separated lines of previous works respectively address i) the prediction of user conversion probability and ii) the attribution of these conversions to advertising events (such as clicks) after the fact. We argue that attribution modeling improves the efficiency of the bidding policy in the context of performance advertising. Firstly we explain the inefficiency of the standard bidding policy with respect to attribution. Secondly we learn and utilize an attribution model in the bidder itself and show how it modifies the average bid after a click. Finally we produce evidence of the effectiveness of the proposed method on both offline and online experiments with data spanning several weeks of real traffic from Criteo, a leader in performance advertising.

Related articles
- Eustache Diemert's slides
- Attribution Modeling Increases Efficiency of Bidding in Display Advertising


November 30 (salle 320 bâtiment UFR IM2AG): Community detection in complex networks by Pierre Borgnat

Complex networks, such as social networks, technological networks, biological or ecological networks or datasets where the relations between data points is relevant, can often be decomposed in communities of nodes (also called modules), that are sub-sets of nodes which are highly connected together and less to the rest of the network. Identifying communities in complex networks is an important issue in network science and is well studied in the past 15 years. We will review several approaches for that, some taking inspiration from computer science, from information  theory, or from physics. We will also present our recent approach, rooted in signal processing, to study in a multiscale manner this issue using graph signal processing.

Related articles
- Community detection algorithms: A comparative analysis
- Graph Wavelets for Multiscale Community Mining

December 14 (salle des séminaires 1, Bâtiment Imag): Neural networks and Random Matrix Theory by Romain Couillet (Ecole Centrale Supélec, Univerty Paris-Saclay)

In this talk, some elementary properties and basics about random matrix theory will be introduced. Some emphasis will be put on the relevance of this theory in the field of statistical learning in high dimension. These  notions will be illustrated on simple neural networks performance analysis. This will shed a new light on the impact of the non linear activation functions used in the network as well as some fundamental limits for elementary model settings in high dimensional spaces.

Related articles
- Application de la théorie des grandes matrices aléatoires à l'apprentissage pour les mégadonnées
- Une Analyse des Méthodes de Projections Aléatoires par la Théorie des Matrices Aléatoires
- A random matrix and concentration inequalities framework for neural networks analysis
- Harnessing neural networks: a random matrix approach
- Signal Processing in Large Systems:A New Paradigm
- Neural networks and random matrix theory (slides)

December 21 (Amphi 22, bâtiment UFR IM2AG): Light propagation in complex media: from imaging, to compressive imaging and machine learning by Sylvain Gigan (The Kastler Brossel laboratory)

Light propagation in complex media is a huge problem for optical imaging:  seeing through a layer of paint, a multimode fiber, or inside biological tissues is a formidable challenge, now made possible thanks in particular to unique tools: spatial light modulators (SLMs), able to digitally encode information on light beams, for display of for wavefront control. This has allowed a revolution in imaging, by allowing focusing light and recovering images at depth in scattering media where all light has been multiply scattered.

I will present our recent work in the domain, with a particular emphasis on our recent experiments, results of a collaboration with signal processing and algorithms experts. We have explored how signal processing, within the recent framework of compressive sensing, can improve imaging using the natural randomness of complex media. More recently, we realized that complex media can be exploited to perform optically a wide range of machine learning tasks. I will present some first proof of principle experiment in image classification.

Related articles
- Measuring the Transmission Matrix in Optics: An Approach to the Study and Control of Light Propagation in Disordered Media
- Imaging With Nature: Compressive Imaging Using a Multiply Scattering Medium

January 18 (salle des séminaires 1, Bâtiment Imag): Big data for telecom by Merouane Debbah (Ecole Centrale Supélec, Univerty Paris-Saclay, Vice President  R&D Center Huawei France, Director of the Mathematical and Algorithmic Sciences Lab)

Mobile cellular networks are becoming increasingly complex to manage while classical deployment/optimization techniques are cost-ineffective and thus seen as stopgaps. This is all the more difficult considering the extreme constraints of 5G networks in terms of data rate (more than 10 Gb/s), massive connectivity (more than 1000000 devices per km2), latency (under 1ms) and energy efficiency (a reduction by a factor of 100 with respect to 4G network). Unfortunately, the development of adequate solutions is severely limited by the scarcity of the actual ressources (energy, bandwidth and space). Recently, the community has turned to a new ressource known as Artificial Intelligence at all layers of the network to exploit the increasing computing power afforded by the improvement in Moore’s law in combination with the availability of huge data in 5G networks. This is an important paradigm shift which considers the increasing data flood/huge number of nodes as an opportunity rather than a curse. In this talk, we will discuss through various examples how the recent advances in big data algorithms can provide an efficient framework for the design of 5G Intelligent Networks.
 
Related articles
- Wireless AI: Challenges and Opportunities
- Machine Learning for Wireless Networks with Artificial Intelligence: A Tutorial on Neural Networks