Sep 13 2017

Melissa Ailem, Shuxia Tang and Georgios Bouloukakis: Awardees of the 2017 Inria@SiliconValley Post-Doctoral Fellowships

Every year Inria@SiliconValley launches a call for Post-Doc Fellowships.

After a tough competition, we are please to congratulate Melissa Ailem, Shuxia Tang and Georgios Bouloukakis who have been selected!

Melissa Ailem Inria@SiliconValley Post-Doctoral Fellow

 

Melissa Ailem will conduct her post-Doc within the LEGO associate team between the MAGNET Inria Team (Lille) and Fei Sha‘s team at USC.

“Transfer and multi-modal learning of word representations”

Recently, word embedding models, e.g., word2vec, have attracted a lot of interest in several research communities, such as machine learning, natural language processing (NLP) and information retrieval. These models seek continuous representations of words that reflect various linguistic regularities between them. Modeling words in such a way has shown great improvements, in various NLP tasks, over the default representation of words as discrete and distinct symbols.

Despite the recent advance in learning distributed representation of words in a vector space, existing models remain far too limited to capture the complexity of text data as they are task-agnostic and fall short of modeling complex structures in languages. The objective is to combine statistical NLP and machine learning techniques to learn finer representations of words. In particular, we will address the following research questions: what are the right representations for structured  data and how  to learn  them automatically? and how to use such representations in subsequent complex and structured prediction tasks in NLP?

To address the above-mentioned challenges, we plan to investigate two main directions. The first direction is in line with the use of multi-modal inputs, e.g., text, images, etc., to learn word embeddings. The second direction is to address the problem of learning good representations for low-resource languages by transferring knowledge from other (high-resource) languages.

 

Shuxia Tang Inria@SiliconValley Post-Doctoral Fellow

 

Shuxia Tang will conduct her post-Doc within the ORESTE associate team between the ACUMES Inria Team (Sophia Antipolis – Méditerranée) and Alexandre Bayen‘s team at UC Berkeley.

“Smart cities: real-time decision making in traffic management”

Due to the complex road topologies and increasing number of vehicles on the roads especially in urban and suburban areas, many traffic problems arise such as traffic jams, car incidents and etc, resulting in large amounts of wasted time and energy. The call for intelligent traffic management and efficient transportation was made long ago, which is a coordinated task among city planners, policy makers and civil engineers. When considering the traffic management problems, researchers generally assume realistically that the city road network is well planned and the transportation policies are perfectly formulated. Under these assumptions, real-time traffic information, collected from various sources such as traffic sensors and video cameras, can be highly counted on. In order to improve the traffic management especially the traffic network flow, an emerging topic is to find intelligent transportation policies based on the traffic data. Although having sufficient information is surely beneficial for formulating on-road decisions, the numerous real-time information supply proposes a major challenge on making the decisions quickly enough or even real-time.

The postdoctoral research plan aims at developing systematic real-time decision-making procedures for the realistic traffic flows on large-scale road networks. In particular, this poses a problem of how to deal with the big data to make real-time decisions. It would be desirable if a systematic control or estimation algorithm could be derived for the class of traffic flow dynamic systems. Many researchers have devoted their efforts to this subject, but most of the results obtained so far are based on discretized PDEs. In the postdoctoral project, we aim to work directly on the continuum traffic flow models of PDE systems, and this would hopefully be one of the first works in this area to the best of our knowledge.

 

Georgios Bouloukakis Inria@SiliconValley Post-Doc Fellow

 

Georgios Bouloukakis will conduct his post-Doc within the MIMOVE Inria Team (Paris) and Nalini Venkatasubramanian’s team at UC Irvine.

“Resilient and Interoperable Interactions for Emergency IoT Applications”

 

The Internet of Things (IoT) promises the integration of the physical world into computer-based systems. Ubiquitous computing devices, featuring sensing capabilities, are deployed in a variety of application domains, such as smart cities, smart factories, resource management, intelligent transportation and healthcare to name a few. However, enacting IoT based systems is still raising tremendous challenges for the supporting infrastructure from the networking up to the application layers. Key challenges relate to deep heterogeneity, high dynamicity, scale, and many others.
IoT devices interact with the physical world primarily by collecting sensing data. This requires access to the data through open APIs and protocols. Over the recent years, IoT devices have been deployed by leveraging several network/communication protocols that enable data access. However, the profusion of such communication protocols introduces technology diversity which results in highly heterogeneous devices. IoT devices penetrate homes, offices, and community spaces and thus, by exploiting their information there are new opportunities to improve peoples’ safety and quality of life. Such devices are used by volunteers to monitor homes/offices for possible indications of seismic activity. Such an application generates critical information and is expected to function correctly and reliably. However, IoT devices can be mobile, low-powered and inexpensive which makes them vulnerable to system changes. Such changes can occur due to a variety of problems including faulty components, inaccurate sensing, intermittent connectivity and software bugs. Resilience is the ability of a system to persistently deliver trustworthy services despite system changes.
An IoT system aimed at handling an emergency situation, like after an earthquake, must be able to: i) dynamically compose the available but probably heterogeneous devices needed for notifying end-users; and ii) ensure the critical data transfer by taking into account system changes and the application’s large traffic flows. Hence, resilient operation of such an application in the presence of failures and disruptions is a key issue. Furthermore, as emergency situations may occur at various scales, such an operation must be equally ensured in large deployments. My post-doctoral research will aim at enabling interoperable, resilient and scalable interactions for emergency IoT applications. This research will be conducted with the Distributed Systems Middleware (DSM) group at University of California Irvine (UCI), USA – in collaboration with the MiMove team at Inria.