Matrix and tensor factorization methods provide a unifying view for a broad spectrum of techniques in machine learning and signal processing, providing both sensible statistical models for datasets as well as efficient computational procedures framed as decomposition algorithms. So far, algebraic or optimization based approaches prevailed for computation of such factorizations. In contrast, the FBIMATRIX project aims to develop the state-of-the-art Markov Chain Monte Carlo (MCMC) methods for Full Bayesian Inference in MATRIX and tensor factorization models. The randomization of Monte Carlo is useful in both Bayesian and non-Bayesian analysis such as model selection, model averaging, privacy preservation or simply better accuracy in computing approximate solutions. MCMC methods are generally perceived as being computationally very demanding and impractical, yet by exploiting parallel and distributed computation, we wish to push the state-of-the-art in terms of scalability, statistical efficiency, computational and communication complexity. In fact, we perceive MCMC as a natural general purpose computational tool of the future for inference and model selection in distributed data, eventually complementing optimization for certain big data problems due to its inherently randomized nature. The project will address Bayesian model selection and model averaging for factorization models, using parallel and distributed computation and current advances in Hybrid Monte Carlo methods that simulate an augmented stochastic dynamics. As such, we aim at developing faster algorithms for hard computational problems such as marginal likelihood estimation and improving convergence rates. We will illustrate the practical utility of the developed parallel and distributed MCMC methods on two challenging applications from two domains: audio source separation and missing link prediction.

The HUSH project investigates the human factors in the production of artificial intelligence (AI) solutions. We look at “micro-workers”—online platform workers who execute essential albeit marginalized data-related micro-tasks which require limited skills, attract low remunerations, and are often paid by piece-rate. These tasks consist, for example, in tagging objects on a photograph to train computer vision models for autonomous vehicles, or in checking the accuracy of transcriptions made by speech-to-text algorithms. Micro-workers are not formally employees but independent contractors or sometimes simple “participants” of the platforms, with varying levels of activity and engagement. While some micro-work is performed through well-known publicly accessible platforms like Amazon Mechanical Turk or Microworkers, tech giants have their own proprietary ones (like UHRS for Microsoft or RaterHub for Google). The factory of the future, construed as a virtual networked infrastructure, puts in place business and communication processes while moving away from the traditional location-based manufacturing paradigm and tipping over into a platform paradigm. As the barriers between outside and inside of the factory are replaced by technological ecosystems where workers are at the core, humans are not vacated from the productive organizations to come—yet their role is often rendered invisible. This project aims to uncover the chains that through platforms, link workers to their clients, French and European companies who demand data-related and algorithmic services. Methodologically, we will leverage a range of economic and social science approaches. The initial phase relies on existing data we have collected: surveys of users of a prominent French platform, a large database of messages from a micro-worker online forum, transaction data from a major international platform, and 92 in-depth interviews. On top of this, a new data collection will be carried out during the project. We will survey 2 000 French SMEs about their usages of micro-work and AI, at three points in time. Furthermore, we intend to collect additional online data through web scraping, API access and extractions via agreements with platforms. Complementary to these quantitative analyses, we will conduct qualitative fieldwork, partly with company managers, union representatives and other local stakeholders (40-50 interviews) and partly with micro-workers, platform operators and business intermediaries in emerging and developing countries where data-related work is outsourced (40-50 interviews). Our study has five main objectives: 1) Study the use of micro-work by firms as a means of outsourcing. Focus is on small and medium sized enterprises (SME) which use intermediaries (platforms) to recruit online labor rather than running their own service. 2) Explore the variety of business models, specializations, and modes of functioning of local and specialized micro-working platforms. We aim to understand their different strategies in terms of pricing, internal governance, and in their degree of transparency and openness to the public. 3) Map cross-country channels through which micro-tasks are outsourced abroad. This will complete and systematize our prior findings towards a robust knowledge base documenting international AI-related outsourcing networks and enabling comparison to spatial patterns previously observed in the English-speaking world. 4) Link business models to working conditions and practices. 5) Develop managerial and policy guidelines, assessing their potential effectiveness in terms of improving the working conditions of individuals active in the micro-tasking platform economy, while enriching reflection around fair and socially responsible AI.

1- Objectives: to design a new air interface for Mobile IoT to be imbedded in future baseline air interfaces for 5G mobile networks and WLAN Our project aims at designing a new air interface for horizon 2020 for high data rate mobile Internet of Things (IoT). It shall support, for instance, new IoT applications such as mobile connected autonomous cameras. The additional cost and energy consumption of the object, due to the connectivity, should remain as low as possible. Compared to the future baseline air interfaces for the mobile networks (5G) and the Wireless Local Area Network (WLAN), this new interface should: support the same mobility with small, low cost and robust devices; achieve similar or slightly lower rate on the device-to-network direction especially; achieve lower energy consumption due to Radio Frequency part; be more sober in terms of radiations. This new interface shall also be embedded in the future baseline interfaces for mobile netwroks (5G) and WLAN standards. It will therefore be compatible with Multi-Carrier Modulations. Indeed, it shall use a pre-defined set of sub-carriers of the future baseline air interfaces for 5G and WLAN without disturbing communications supported by the other sub-carriers of the same air interfaces. 2- Scientifical approach and technical challenges: spatial modulation and tunable locally resonant antennas To this aim, we will design a new wireless Transmitter/Receiver to be installed on the connected mobile objects, and we will jointly optimize the use of two disruptive techniques: Spatial Modulation Multiple Input and Multiple Output (SM-MIMO) and Tunable Locally Resonnant Antennas (TLRA). On the one hand, compared to a conventional MIMO system, SM-MIMO theoretically reduces the cost in complexity and energy of the object, per transmit data bit per second, by using a Single Radio Frequency chain instead of several. On the other hand, TLRA theoretically increases SM-MIMO performance thanks to its larger flexibility. One can therefore hope that by jointly optimizing both techniques, the goals of the projects will be met. 3- Work program: from design to proof of concept The first main output of the project will be a material proof of concept with over-the-air data transmission and using TLRA antennas will be issued. The second main output of the project will be a set of innovative cross PHY-Antenna schemes relying on SM-MIMO and TRLA, to address the challenging requirements of mobile networks: high mobility, Frequency Division Duplex mode and compatibility with Multi-Carrier based waveforms. Metrics such as the cost in energy, complexity and radiation per transmitted bit per second will be assessed. 4- Consortium: mainly academics, leaded by industrials, and experts in cross PHY-Antenna design otpiimsation The consortium is composed of 2 companies (Orange and TRCOM) and 4 academic institutions (Langevin Institute, IETR INSA of Rennes, Telecom Bretagne, Supélec). Orange will lead the project. Supélec is leading the global research on SM-MIMO topic. The other partners are producing together a proof of concept of joint Time Reversal and Micro structured antennas in the ANR TRIMARAN project (ending in June 2014), and are studying a basic Receive Spatial Modulation scheme. This consortium has therefore the following advantages to achieve the project proposal objectives: there is a strong complementary between partners; there is an opportunity for savings in material and FPGA development resources by reusing TRIMARAN project platforms, to address new challenges; the consortium has experience on joint PHY-Antennas design for small equipments.