Abstract
From LRDE
H
Decision Diagrams are a family of data structures that represents huge data sets using a small amount of memory. These structures can be of fixed size (tuples), or varying size (lists, maps, ...), the DD handling being different for each one. Data Decision Diagrams and Set Decision Diagrams handle varying size data thanks to operations named homomorphisms. However, the definition of a correct operation can be hard because numerous errors hard to identify can happen. This seminar offers a presentation of an algorithms library that gives a more abstract view on handled data. This library contains the algorithms defined in "List" and "Map" modules from the Objective Caml standard library, allowing the user to focus on his specific problem. +
Les Diagrammes de Décision sont une famille de structures de données permettant de représenter avec peu de mémoire de grands ensembles de données. Ces structures peuvent être de taille fixe (un tuple) ou variable (une liste, un conteneur associatif, ldots), la manipulation du DD ne se faisant pas de la même manière. Les Data Decision Diagrams et Set Decision Diagrams manipulent des données de taille variable grâce à des opérations, les homomorphismes. Cependant la définition d'une opération correcte peut dérouter l'utilisateur, et passe souvent par de nombreuses erreursdifficiles identifier. Ce séminaire propose une bibliothèque dùalgorithmes fournissant une vue plus abstraite que les homomorphismes "bruts" des données manipulées, en reprenant les algorithmes définis dans les modules "List" et "Map" d'Objective Caml. L'utilisateur peut se concentrer sur les parties spécifiques à son problème. +
I
Actually, i-vector space is the most recurrent representation of speech information in speaker recognition systems. The scoring process is generally based on Cosine Distance (CD) or Probablistic Linear Distriminant Analysis (PLDA) methods. The aim of this work is to replace these approaches by a MultiLayer? Perceptron (MLP). The MLP showed good performances in nonlinear function approximation. The main idea is to find better functions than cosine scoring method. The performance of the MLP method will be compared with other methods such as CD, PLDA and Restricted Boltzmann Machines (RBM) method presented by Jean Luc. +
In this work we apply Convolutional Neural Networks to the task of speaker recognition. The CNN is used to approximate a measure of the distance between two i-vectors (representation of a speaker). Contrary to the commonly used cosine similarity measure, the function approximated by a CNN can be non-linear. The performance of this model will be compared to the ones of the Cosine Similarity measure and PLDA classification. +
Building a robust speaker recognition system needs very slow and complex algorithms. In this work, we propose a deep neural network to map a low dimensionnal ivector space based on less complex model into high dimensional ivector space. The system will be evaluated on speaker recognition task of NIST-SRE 2010 data. +
Ce rapport présente l'implémentation en C++ de l'algorithme ID décrit par Dana Angluin et al dans Polynomial Identification of omega-Automata. Il permet l'identification, ou l'apprentissage passifd'omega-langages réguliers et des omega-automates associés, dans un temps et une mémoire polynomiaux. C'est un travail préliminaire à l'étude de l'apprentissage actif d'omega-langages. Le code est disponible sur https://gitlab.lrde.epita.fr/cpape/ID +
Les botnets sont devenus la technique principale pour effectuer des attaques au sein d'un réseau. Ils ont été utilisés dans le passé pour voler des informations auprès d'une organisation ou envoyer des spams. Plus récemment, les attaquants s'en sont servis à des fins financières pour par exemple miner des bitcoins. Dans cette perspective, il est essentiel de détecter ces réseaux malicieux afin de défendre les intérêts des utilisateurs ou d'organisations. La recherche publique a souvent 'dté derrière la forte adaptation des attaquants aux systèmes de détections d'intrusions. Nos travaux consistent en l'utilisation de techniques d'apprentissage automatiques non-supervisées encore inexplorées dans ce domaine afin de détecter différents types de botnets. +
Botnets are the primary way of attacking computer networks and are being used to steal information, spy organizations or send spams, by compromising devices connected to the internet. More recently, botnets have also seen themselves being used for financial interests such as mining bitcoins at a large scale. It is a primary threat which is essential to identify in order to defend the interests of users and any type of organization. Unfortunately, public research has often been one step behind the fast adaptation of attackers to detection systems. Our work consists in using unsupervised machine learning techniques unprecedentedly used on such tasks to detect botnets on different scenarios. +
Morphological operators can be applied on grayscale images to filter out some objects or conversely to emphasize some objects. Therefore, by choosing an appropriate structuring element, it is possible to eliminate some elements of a map like text and to reconstruct discontinuous lines. A filter bench has been defined to extract from images of ancient maps the thin lines whatever their orientation, that isthe plot boundaries. On the resulting images, a watershed algorithm can afterwards segment the plots. Moreoverapplying a "seam carving" algorithm as pre-processing removes grid lines when needed. +
As part of a partnership with the Gustave Roussy Institutthe LRDE's image processing library Milena, offers an application dedicated to image reconstruction. Different images of the same object but obtained from different modalities have to be processed. First, these images are simplified. Then objects contained by these images are extracted. The final step is to mix information into a unique image. Thereby, the process is composed of several stages: image filtering, segmentation, binarizationmultimodal image registration and image reconstruction. The presentation will especially focus on the segmentation part. +
Milena is the generic image processing library of the Olena platform. The library aims at remaining simple while providing high performances. The introduction of new image types based on graphs has revealed some design problems limit ing its genericity. For instance, we have always considered that "images have points"; yet some images have sites that are not points (but edges, facets, and even sets of points). Another erroneous assumption was to consider that sites are localized by a vector (e.g., (x,y) in the 2D plane), which cannot be true when sites are not point-wise. Therefore there was a need to reconsider the image types and their underlying images properties.In this seminar, we will present a new image taxonomy that solves those issues. +
FIXME +
FIXME +
In the model checking field, Partial Order Reduction (POR) is a method which allows to notably reduce the size of datastructures used to represent the different possible executions of program model, by considering only representative execution paths, insteed of all. This has a cost: information is lost. When the goal is only to check the absence of deadlocks, enough information is kept, so POR works well. But when in comes to LTL model checkingrelevent information may be dropped: execution paths which modifiy the value of AP (Atomic Proposition) occuring in the LTL formula. Moreover, it is only possible to use POR for LTL X. Invisible and transparent transitions are methods which add additional conditions to fulfill during the POR, in order to keep all execution paths which may modify the value of the AP. We explain these methods and how they have been implemented in the Spot model checking library. +
La morphologie mathématique est devenu un outil indispensable pour une bibliotheque de traitement d'image. Les algorithmes produit grace a cette technique sont tres efficaces et permettent des résultats tres satisfaisant notamment pour des opérations de segmentation d'image. Le travail produit a donc pour objectif d'implémenter et d'intégrer dans la bibliotheque Pylene les représentations hiérarchique morphologique. Notre travail s'est concentré principalement sur la pipeline globale de segmentation utilisant ces méthodes, ainsi que la visualisation grace aux cartes de saillance. L'objectif etant bien entendu de réaliser les algorithmes les plus optimisés possible pour pouvoir les utiliser sur de grandes images. +
The Common Lisp language provides a predicate functionSUBTYPEP, for instrospecting sub-type relationship. In some situations, and given the type system of this languageknowing whether a type is a sub-type of another would require enumerating all the element of that type, possibly an infinite number of them. Because of that, SUBTYPEP is allowed to return the two values (NIL NIL), indicating that it couldn't answer the question. Common Lisp implementations have a tendency to frequently not answereven in situations where they could. Such an abusive behavior prevents potential optimizations to occur, or even leads to violating the standard. In his paper entitled “A Decision Procedure for Common Lisp's SUBTYPEP Predicate”Henry Baker proposes an algorithm that he claims to be both more accurate and more efficient than the average SUBTYPEP implementation. We present here the current state the current state of our implementation and discuss one potential improvement based on R-trees of Baker's algorithm. +