Imagine you have to answer the following questions: how to build a computer-aided diagnosis tool for neurological disorders from images acquired from different medical imaging devices? that could identify which emotion is feeling a person from her face and her voice? How could these tools be still operational even when some data of a type is missing and/or poor quality? These questions are at the core of some problems addressed by the Institut de Neurosciences de la Timone (INT), where people have expertise in brain imaging based medical diagnosis, and Picxel, a SME centered on affective computing. The Laboratoire d'Informatique de Paris 6 (LIP6), the Laboration Hubert Curien (LaHC), and the Laboratoire d'Informatique Fondamentale de Marseille (LIF, head of the PI) are the other partners that are teaming up with INT and Picxel: in this project, they provide their renowned knowledge in machine learning, wherein they have developed, theoretical, algorithmic, and practical contributions. The five partners will closely work together to propose original and innovative advances in machine learning with a constant concern to articulate theoretical and applicative findings. The above questions pose the problem of (a) building a classifier capable of predicting the class (i.e. a diagnosis, or an emotion) of some object, (b) that of taking advantage of the few modalities or *views* used to depict the objects to classify and, possibly (c) that of building relevant representations that take advantages of these views. This is precisely what the present project aims at: the development of a well-founded machine learning framework for learning in the presence of what we have dubbed *interacting views*, and which is *the* notion we will take time to uncover and formalize. To address the issues of multiview learning, we propose to structure as follows. On the one hand, we will devote time to establish when and how classical (i.e. monoview-based) learnability results carry over to the multiview setting (WP1); this may require us to brush up on our understanding of different notions and accompanying measures of interacting views. On the other hand, possibly building upon the results just mentioned, we will build new dedicated multiview learning algorithms, according to the following lines of research: a) we will investigate the problem of learning (compact) multiview representations (WP2), then b) we will create new algorithms by leveraging some recent works on transfer learning -- multitasks and domain adaptation -- to the multiview setting (WP3), and, c) we will address the scalability of our algorithms to real-life conditions, such as large-dimension datasets and missing views (WP4). Finally, the performances of our learning algorithms will be assessed on benchmark datasets, both synthetic and real, that we will collect and make available for the machine learning community (WP5). Beyond the mere evaluation of our algorithms, these dataset will be disseminated to promote reproducible research, to identify the most suitable algorithms in a multiview setting, and to make the machine learning community aware of the exciting problems of multiview learning for affective computing and brain-image analysis.

There is an emerging trend in robotics : rather than by a few bulky robots, certain tasks can be performed by tiny robots, each with very limited capabilities. This trend is similar to the less recent trend about sensor networks, with sensors that are also of limited capabilities but deployed in huge number. This is the emergence of systems where elementary bricks are simple and cheap though can provide relatively complex collective behavior. From an algorithmic point of view, one need to consider a new computing paradigm « Moving and computing »: the study and design of systems where the computational entities themselves are capable of movement within the spatial universe they inhabit. The field has applications in areas as diverse as autonomous robots moving in a terrain, software agents moving in a network, autonomous intelligent vehicles, wireless mobile ad-hoc networks, and networks of mobile sensors; where the computational objectives are exploration, coordination and cooperation. When considering the design of algorithms for mobile robots in a geometric environment, the modeling of the environment, i.e., the way the mobile entities have access to it, is crucial. The entities can have access to only limited aspects of the environment : e.g. where it can move. Indeed, in this setting, the robot main responsibility is to compute where to go next. Such a modeling implies the study of the graph of the possible locations, linked by the elementary moves. Such a graph does not have an arbitrary structure but inherit some combinatorial properties from its geometric context. In this framework, using the mobile agent model, from classical distributed computing, is very much relevant. The specificity of our study is that the graphs under consideration are of geometric type (e.g. the visibility graph of a polygon). Moreover, it is known that adding more sensing capabilities will yield more efficient algorithms. A natural investigation is therefore to characterize what are the weakest kind of sensors, i.e., the kind of geometric information, that enable to solve efficiently problems such as exploration, map construction, rendezvous, ... In contrast with the previous situation, in the case of mobile sensors, the computation is more how to react to moves and changes in the topology that are not directly under control. However, similarly, by using geometric properties, it is possible to improve the efficiency of algorithms, compare to algorithms using only combinatorial graphs properties. Hence, solving tasks such as broadcast, computing/repairing local structures, benefits from a deep understanding of the relationship generated by the actual topology of the sensor network, especially the possible dynamic evolutions of such networks. It is also possible to show that the ``mobile agent" model is also relevant in this context, because it enables to use a simple computational structure within a complex data structure. A distant goal for our research, that will be of importance in the near future, is to investigate thoroughly the interactions and evolutions of mobile agents in dynamic networks. More generally, our objective is to dramatically improve the models and algorithms for distributed robot computing. Such a study implies to get a better understanding of the relations between local, global and geometric properties shared by those problems and environments. We will benefit from tools and known results from geometric graph theory, discrete algebraic topology, computational geometry and timed systems. Last, part of the validation of this project's results will be done with « LED's CHAT » (http://leds-chat.net/): a modular light and sensor system developed at LIF, and currently subject to a technology transfer program with the aim of commercializing the technology to actual light applications.

The main goal of the Aggreg project is to develop efficient algorithms for answering aggregate queries for databases and data streams of various kinds. Aggregate queries are central for computing statistics on data collections: Rdf stores, NoSql databases, streams of data trees in Xml or Json format, uncertain databases, relational databases, and datawarehouses. Considering that counting is the basis of aggregate queries the principal difficulty here is to overcome the inherent computational hardness of many counting problems, which precludes general and efficient solutions. Instead, we propose to: study the complexity of expressive fragments of the class of aggregate queries, search for efficient algorithms on tractable fragments, identify which parameters can be fixed in order to obtain tractability, find general algorithms that are gracefully degrading, and also efficient approximation algorithms. We apply methods from algebra, automata, probability, algorithmics and complexity theory.

The audio inpainting concept, recently proposed by the coordinator and colleagues, is a conceptual breakthrough that unifies in a single framework all the audio processing problems where data is partially missing or highly degraded. Instances of such problems are click removal, CD scratches restoration, declipping, packet loss concealment, source reconstruction in the time-frequency domain and bandwidth extension. While these tasks had been addressed separately in the past, the audio inpainting unified formulation as an inverse problem is a promising abstraction to factorize the main difficulties shared among tasks, to provide methods that outperform state-of-the-art techniques on existing tasks and to address new problems where missing data reconstruction has been too difficult a task so far. The MAD proposal develops audio inpainting for any task involving missing audio data. The main objectives of this proposal are: a) to deploy the concept of audio inpainting within the research community by proposing new approaches, by addressing new applications and by creating and animating a dedicated research network; b) to initiate works on time-frequency inpainting, i.e. on the reconstruction of missing coefficients in a transform domain; c) to expand the concept of and the techniques for audio inpainting by developing connections with machine learning. The project establishes strong relations between signal processing and machine learning. It does not only consist in applying machine learning techniques to signals but also deals with a machine learning formulation of signal processing problems and with the integration of computational trade-offs in algorithms. The project also draws connections between audio and image processing. The proposal implies close interactions between theory and applications with top/down and bottom/up relations. All those original aspects are revealed in the composition of the team and are expected to result in powerful approaches to real applications. The MAD proposal is submitted to the ANR JCJC program under the leadership of Valentin Emiya, this proposal being the largest project he coordinates. To address its ambitious and diverse objectives, MAD involves a large team of 11 members, with research experiences in theory and application views from both academy and industry, signal processing and machine learning. Seven team members are located in the same site, the remaining four members being in three isolated distant sites including two members at Technicolor.