Thursday 29 November 2012
User driven innovation at CEA Tech
Speaker: Timothée Jobert
, researcher at CEA-Leti
, Grenoble, France
This talk gives a tour d'horizon of the CEA Technological Research Department and focuses on the open innovation platform. For the last 10 years we have been seeking to mix technological sciences with research in art/creativity and social sciences. Through an incremental approach of collaboration with technology project teams, over the years we have reached an efficiency that reinforces our position and legitimacy inside the company. This is witnessed by the evolution of our position in the technology teams: from concept and prototype assessments to co-design / from outsider (service provider) to insider (team member). We shall illustrate this evolution through several examples in the fields of mobility (bicycle, electric car, car sharing) and energy (smart grids, smart city, air quality management etc).
Tuesday 27 November 2012
Boosting k-NN for large scale image classification
Speaker: Michel Barlaud, professeur émérite, membre senior de l'Institut Universitaire de France at Université de Nice-Sophia Antipolis
, Nice, France
Recent works show that large scale image classification problems rule out computationally demanding methods. On such problems, simple approaches like k-NN are affordable contenders, still with room space for statistical improvements. We first present algorithm UNN, showing how to leverage k-NN to yield a formal boosting algorithm. Second, we propose N3, an Adaptive Newton-Raphson scheme to leverage k-NN. We show that it is a boosting algorithm, with several key algorithmic and statistical properties. N3 brings efficient estimators of posteriors at no additional cost. We propose a divide and conquer algorithm in order to cope with the classical curse of dimensionality of NN search. Experiments are provided on the SUN, Caltech and ImageNet databases. They confirm that boosting a subsample — sometimes containing few examples only — is sufficient to reach the convergence regime of N3. Under such conditions, N3 challenges the accuracy of much more computationally demanding contenders.
Thursday 22 November 2012
Online reputation monitoring & conversation analysis at Websays: the business and the science
Speaker: Hugo Zaragoza, chief executive officer and founder at Websays
, Barcelona, Spain
Websays strives to provide the best possible analysis of online conversation to marketing and social media analysts. One the of the obsessions of Websays is to provide "near-man-made" data quality at marginal costs. I will discuss how we approach this problem using innovative machine learning and UI approaches.
Thursday 15 November 2012
Contributions to the study of paraphrases and their applications in NLP
, associate professor at Université Paris Sud
, Orsay, France
Natural language allows for a large variety of expressions to formulate a given content. One of the major challenges that Natural Language Processing applications face is the ability to recognize and/or appropriately generate paraphrases. We will present an overview of such applications, and describe some of the work conducted at LIMSI-CNRS on the acquisition of paraphrase from a variety of corpus types. In a second part, we will present in some details our work in two areas: assisting writers during text revision, and automatically generating alternative expressions to improve the translatability of a given text. We will conclude by describing some ongoing and future work, focusing on the difficult problems that we have identified.
Tuesday 13 November 2012
Block-coordinate Frank-Wolfe optimization with applications to structured prediction
Speaker: Martin Jaggi
, post-doctoral researcher at Ecole Polytechnique
, Palaiseau, France
Coordinate descent on one hand, and the Frank-Wolfe algorithm on the other hand are two of the earliest known first-order methods for convex optimization. Here we will combine the two methods to obtain a new randomized block-coordinate optimization algorithm for block-separable constrained problems, which appear for example in machine learning and computer vision. This is motivated by the hope to combine the advantages of the two methods, namely the cheap iteration complexity of coordinate descent, and the sparse iterates and primal-dual convergence guarantees from Frank-Wolfe. Read more
Thursday 08 November 2012
Models of input dependent covariances between multiple responses
, doctoral candidate at University of Cambridge
, Cambridge, U.K
Accounting for input dependent covariances between multiple responses can greatly improve statistical inferences. For example, if we wish to predict the expression level of a gene (response) at a particular time (input), it helps to consider the expression levels of correlated genes, and how these correlations depend on time. I will discuss three new models I have introduced for input dependent covariances: the Gaussian process regression network (Wilson et. al, 2012), generalised Wishart processes (Wilson and Ghahramani 2011), and copula processes (Wilson and Ghahramani, 2010). I will describe the connections between these models, and the high level ideas that can be applied to regression and classification in general, to improve predictive performance. I apply these models to problems in econometrics, geostatistics, and large-scale gene expression. The models prove to be scalable with greatly enhanced predictive performance over alternatives: the extra structure being learned is an important part of a wide range of real data.
Thursday 25 October 2012
Semantic attributes for object categorization
Speaker: Christoph Lampert
, researcher at Institute of Science and Technology
, Klosterneuburg, Austria
Attribute-based categorization is the task of classifying images based on a description of the target class. Because it does not rely on training examples to build a class model, it allows tackling seemingly impossible tasks, such as zero-shot learning. In the talk I will give an overview into the working mechanisms of attribute-based categorization, in particular highlighting its use for fine-grained categorization, where classes are visually similar and training examples are rare, but discriminative knowledge by human experts is often available.
Thursday 18 October 2012
Rental of a durable good
Speaker: Jacques Crémer
, director of research at University of Toulouse I
, Toulouse, France
Coauthors: Cyril Hariton, European Commission, and Sinem Hidir, Toulouse School of Economics.
How should the owner of a durable good rent it to agents who desire to use it for different lengths of time? This question is important for many network industries: there are short run and long run users of gaz pipelines, and airports must choose between giving a particular slot to a regular airline or to keep it open for irregular charter flights. In order to examine this question, we build an infinite horizon stationary model where a monopoly seller rents a good. At each period, a number of potential buyers appear, with different lengths of demand. We study and compare the mechanisms that would be used by a profit maximizing and by a social welfare maximizing seller. We show that, in some precise sense of the term, a profit maximizing seller will favour long term renters.
Thursday 11 October 2012
Current Investigations on Statistical Machine Translation at LIG (Laboratoire d’Informatique de Grenoble)
Speaker: Laurent Besacier
, professor at Université Joseph Fourier
, Grenoble, France
This presentation will focus on two topics currently addressed in Statistical Machine Translation at LIG.
The first one is related to Confidence Estimation at word level. We investigate various types of features to circumvent this problem, including lexical, syntactic, semantic and system-based features. Our experimental results using a classifier based on conditional random fields (CRF) will be presented as well as a "Feature Selection" strategy applied to better understand the role of each type of feature. The use of Boosting to combine multiple sub-models and take advantage of their complementarity will be also presented. Read more
Wednesday 26 September 2012
Evaluating social applications
Speaker: Sihem Amer-Yahia, directrice de recherche at CNRS, Université de Grenoble
, Grenoble, France
Social applications have been receiving increasing attention from both industry and academia. A social application is characterized by the addition of value to a context where users interact with each other and with existing resources, tagging them, rating them or simply browsing them. Traditionally, applications on the Web have been evaluated using a multitude of methods ranging from the least intrusive ones such as offline A/B testing to the most intrusive such as online bucket testing. The same methods have been applied to evaluate social applications. I will describe an example of a social application and its evaluation using crowdsourcing and I will end the talk with a discussion on the need to better understand and formalize the evaluation of social applications.