XRCE organises public scientific seminars on a regular basis which you are welcome to attend. These seminars are an occasion to exchange with researchers from various backgrounds and to broaden scientific expertise. You can subscribe to our seminar RSS Feed for dates, speakers and topics.
We address the problem of planning (sequential decision making) under uncertainty, modeled as a discrete-time optimal control problem. The solution of this problem is a control policy that maximizes expected future reward. We briefly review Markov decision processes (MDP) and dynamic programming. We then discuss the partially observable Markov decision process (POMDP) where the system state is stochastically coupled to an observation signal (a POMDP is a parametrized HMM with rewards). Planning in a POMDP requires reasoning over a continuous high (or infinite) dimensional belief space, which renders exact dynamic programming methods intractable. We describe some approximate dynamic programming methods for POMDPs that offer good performance trade-offs. Finally we briefly discuss the case of model-free optimal control (reinforcement learning), and a recent result that allows casting the optimal control problem as a maximum likelihood problem.Slides (1.28 MB)
The LIPN has developed shallow but robust sentiment analysis techniques to a number of domains including financial news, blog posts and consumer-generated media in general. In the framework of Infomagic, we are investigating the benefits of using such an approach to spot and ultimately extract key textual elements that indicate potential crisis in a broad sense. On one hand, the approach we present aims at providing a rough global and local classification of texts and spot phrasal cues for crisis; used efficiently, this information can save work to systems down the pipeline by pruning irrelevant segments of text. On the other hand, this approach can serve as a fall-back to deeper (linguistic) analysis, especially when data are particularly noisy or irregular. The seminar will explain the approach and demonstrate its use to the analysis of online short news.
Work done in collaboration with Michel Généreux.Slides (1) (282.26 kB)
The field of Structured Learning extends the idea of SVMs, allowing us to deal with structured output spaces (such as graphs or images). In this talk, I will present a brief introduction to structured learning, and demonstrate how it can be used in applications such as image classification.Slides (2.28 MB)
Building a memory of the large mass of information that only exists on the Internet, specifically on the web, is a necessity and a challenge. A necessity as otherwise, a large amount of traces of today's society will be lost; a challenge because of the scale and variety of techniques of publications that can be found on the web.
We will review these challenges as well as opportunities that building such corpus bring for research, introducing the possibility to work on trends and evolutions instead of frozen snapshots. We will specifically present two recent European projects in this domain, Living Web Archives (FP7 IST) and Living Knowledge (FP7 FET).Slides (2.45 MB)
We will discuss some Matching Pursuit (MP) algorithms in machine learning called Kernel Matching Pursuit (KMP) and Matching Pursuit Kernel PCA (MPKPCA) and derive generalisation error bounds that upper bound their loss. The MPKPCA bound turns out to be a compression scheme and can be bounded using sample compression theory. We show that this bound is tighter than previously proposed bounds for KPCA. The KMP bound is less non-trivial and we make the application of compression bounds to VC theory -- hence viewing the KMP algorithm as a "compression scheme" in feature space.
The MP algorithms can be described as 2-step algorithms that 1) maximise some function, and 2) carry out deflation in order to construct learners that achieve dual sparsity. To this end we will discuss extensions of MP to Kernel Canonical Correlation Analysis (KCCA) and Kernel Fisher Discriminant Analysis (KFDA) (time permitting), as well as there respective generalisation error bounds.Slides (267.02 kB)