CSI Seminar 2016-07-01
10h00 Domain Mismatch Compensation for Text-Independant Speaker Recognition – Valentin Iovene
The impact of domain mismatch when the system training data and the evaluation data are collected from different sources remains a challenge. This study lays out state-of-the-art techniques used for domain mismatch compensation such as a library of whitening transforms and the use of a dataset-invariant covariance normalization matrix to obtain domain-invariant representations of feature vectors.
10h30 Bottleneck neural networks for Speaker Recognition – Guillaume Daval-Frerot
Deep neural networks are increasingly used for their capacity to correlate concrete parameters to deduce abstract characteristics. The bottleneck neural network is a specific form of those. This work presents the principle of this kind of network and its use for the reprocessing of Mel Frequency Cepstral Coefficients in a speaker recognition system. Therefore, it is a matter of studying the convergence of such network but also the change in overall system performance.
11h00 Metric Learning using a Siamese Deep Neural Network – Anatole Moreau
This work uses a siamese architecture to learn a similarity measure. We apply two different samples over two identical sub-networks with the same set of weights. The input of each network is based on statistical information of speech data. We can then compute the distance between both informations. The DNN is able to reduce the dimensionality of the input because it learns an invariant mapping. We present the results of the learned similarity metric using different kind of informations and compare them to classic metrics based on PLDA or cosine similarity applied to i-vectors.
13h00 Alternating automata support – Amaury Fauchille
Alternating automata add a universal branching to the existential branching from non-deterministic automata. Büchi alternating automata are exponentially more concise than non-deterministic Büchi automata. Additionally, they are well suited to translate linear temporal logic as their size is proportional to the translated formula. This work aims to add alternating automata support to Spot. This will make Spot compatible with other tools producing or using aternating automata, and allow future algorithms working on alternating automata to be implemented.
13h30 Product of Parity Automata – Laurent Xu
Spot, a library for transition based Spot, a library for transition-based '"`UNIQ--math-00000039-QINU`"'-automata manipulation, provides a method to determinize Büchi automata which can produce quasi-parity automata. We introduce three tools to manipulate these automata which are reducing the number of sets in the acceptance condition, changing the style of the parity acceptance expression and colorizing an automaton. Olivier Carton introduced a method to construct the product of state-based parity automata which keeps the parity acceptance. We adapt this method to make it work with transition-based automata and we present optimizations of this product. Our work gives a better support of automata with parity acceptance and may also lead to further optimizations of the determinization of Büchi automata.
14h00 Automatic segmentation of Cassini's maps – Anne-Claire Berthet
Our goal is to segment ancient maps of France, being cut and glued on a canvas. The colors in the canvas and in the maps are very close, making the boundary difficult to distinguish. However, pixel-perfect parts of the boundary are obtained thanks to morphological operators. To complete the segmentation, the image is separated, thanks to a maskinto three zones (map, canvas, and a "thick" boundary)which are used by a classification method to obtain results close to the actual boundary. Finally, the first results are corrected thanks to such new knowledge.
14h30 Supervised Discrimination of Characters on Images – Thibault Deutsch
The discrimination of characters is an important domain of optical characters recognition. The goal is to determine if a delimited surface of an image is a character or not, with rotation invariance. We are able to reduce the redundant information by doing a principal component analysis (PCA) on the training data set. Then, we use the probabilistic linear discriminant analysis (PLDA) algorithm to models both intra-class and inter-class variance as mutli-dimensional Gaussians. The performance of the new model will be compared with the one currently used in the optical characters recognition application of Olena.
15h30 Random rational expression generation – Lucien Boillod
In this report, we present the implementation of an efficient and generic algorithm to generate random weighted rational expressions, including multitape rational expressions. It support any operators, labels and weights present in Vcsn. This tool allows a better coverage for the tests. We also present a way to generate random paths on weighted automata.
16h00 Quotient of weighted automata and rational series – Thibaud Michaud
In this report, we show how the quotient operator has been implemented in Vcsn, a generic and perfomant automata manipulation library. After defining the left and right quotient over rational series, we explain the algorithm implemented in Vcsn to compute the quotient of two automata. We then explore the consequences of introducing the operator on the expression-side of the library, and particularly on expansions.
16h30 K shortest-paths in Vcsn – Sébastien Piat
When trying to retrieve multiple shortest paths in a graph, we cannot simply run a shortest path algorithm multiple times as it would always retrieve the same. Some algorithms exist to solve this problem. One of them, the Yen algorithm hides transitions in the graph between each calculation to retrieve the correct paths. This work will present how it was implemented in Vcsn and the optimization techniques we tried on this algorithm (including the implementation of a sparse set).
17h00 Contribution to dyn:: – Raoul Billion
C++ provides a good support for performance and genericity through templates. However, it has a real cost on flexibility and compilation time. In order to overcome these issues, Vcsn introduces a dynamically-typed library named dyn. The purpose is to bring more flexibility to the project and it also allows just-in-time compilation. This work presents, first of all, our thoughts on how to handle and optimize our compile time, then various improvements added to dyn.