Salem Benferhat (CRIL, Université d’Artois - CNRS)
This short talk addresses the problem of handing conflicts in lightweight ontologies. Recently, eight genuine ways to deal with conflicting information have been identified in the context of lightweight ontologies. This paper focuses on a so-called IAR semantics and proposes some desirable properties and algorithms to extend it to prioritized Aboxs.
Zied Bouraoui (CRIL, Université d’Artois - CNRS)
Word embeddings have been found to capture a surprisingly rich amount of syntactic and semantic knowledge. However, it is not yet sufficiently well-understood how the relational knowledge that is implicitly encoded in word embeddings can be extracted in a reliable way. In this talk, I will present two probabilistic models that address this issue. The first model is based on the common relations-as-translations view, but is cast in a probabilistic setting. Our second model is based on the much weaker assumption that there is a linear relationship between the vector representations of related words. Compared to existing approaches, our models lead to more accurate predictions, and they are more explicit about what can and cannot be extracted from the word embedding.
Séminaire Un tour d'horizon des bandits
Frederic Koriche (CRIL, Université d’Artois - CNRS)
Imaginez le jeu répétitif suivant : à chaque tour, vous devez choisir une “action” parmi N décisions possibles et, une fois l’action choisie, l’environnement vous répond par une “récompense”. L’objectif est de minimiser le regret entre vos choix effectués et la meilleure action que vous auriez sélectionnée, si vous aviez connu à l’avance les récompenses données par l’environnement. Ce cadre, appelé “bandits multi-bras”, est au coeur de l’apprentissage par renforcement et apparait dans de nombreuses applications en IA. L’objectif de ce tutoriel est de donner un panorama des principaux problèmes et algorithmes de bandits. Les problèmes sont regroupés en classes, selon le type d’environnement (stochastique vs adversarial), le type d’action (simple vs combinatoire) et le type de renforcement (informatif, semi-bandit, bandit). Nous présenterons quelques algorithmes de bandits pour ces diverses classes, et illustrerons l’exposé par quelques applications en IA.
Séminaire Dependency weighted aggregation
Florent Capelli (Université de Lille)
It is well-known that one may faithfully approximate the
expected value of a random variable by taking the average of many
independant observations of the variables. The average quickly
converges to the expected value with good probability. This fact is
commonly used when computing statistics in data mining.
However, in this setting, the assumption that the observations are
independant is usually wrong and we have no more guarantees on the
convergence of the average toward the expected value. One way of
circumventing this problem would be to only use an independant subset
of the observations. Finding a large independant subset is however a
hard problem and this method tends to discard precious data.
An other method proposed by Wang and Ramon starts by assigning weights
to each observation. If the weights are solution to some linear
program, then we have the guarantee that the weighted average
converges toward the expected value.
In practice however, when the number of factor of dependencies is
large, the size of such a linear program can be hudge. In this talk,
we will present a way of generating an equivalent compressed linear
program when the dependencies are succinctly represented.
This is joint work with Nicolas Crosetti, Jan Ramon and Joachim
Séminaire Learning Conceptual Spaces from Data
Steven Schockaert (Cardiff University)
Conceptual spaces have been proposed by Gärdenfors as an intermediate knowledge representation framework, sitting between high-level symbolic representations and low-level neural representations. While the theory of conceptual spaces has been influential in philosophy and cognitive science, to date it has seen relatively few applications within the field of Artificial Intelligence. One of the main stumbling blocks for such applications is that learning conceptual space representations in a purely data-driven way is challenging. In this talk, I will first give an overview of some approaches for learning conceptual space representations that have been developed within the context of the FLEXILOG project.