Vincent Guigue

BCI Competition III: Dataset II- Ensemble of SVMs for BCI P300 Speller

Brain-computer interface P300 speller aims at helping patients unable to activate muscles to spell words by means of their brain signal activities. Associated to this BCI paradigm, there is the problem of classifying electroencephalogram signals related to responses to some visual stimuli. This paper addresses the problem of signal responses variability within a single subject in such brain-computer interface. We propose a method that copes with such variabilities through an ensemble of classifiers approach. Each classifier is composed of a linear support vector machine trained on a small part of the available data and for which a channel selection procedure has been performed. Performances of our algorithm have been evaluated on dataset II of the BCI Competition III and has yielded the best performance of the competition.

Ensemble of SVMs for improving brain computer interface p300 speller performances

This paper addresses the problem of signal responses variability within a single subject in P300 speller Brain-Computer Interfaces. We propose here a method to cope with these variabilities by considering a single learner for each acquisition session. Each learner consists of a channel selection procedure and a classifier. Our algorithm has been benchmarked with the data and the results of the BCI 2003 competition dataset and we clearly show that our approach yields to state-of-the art results.

Pedestrian detection using stereo-vision and graph kernels

This paper presents a method for pedestrian detection with stereovision and graph comparison. Images are segmented thanks to the NCut method applied on a single image, and the disparity is computed from a pair of images. This segmentation enables us to keep only shapes of potential obstacles, by eliminating the background. The comparison between two graphs is accomplished with an inner product for graph, and then the recognition stage is performed learning is done among several pedestrian and non-pedestrian graphs with SVM method. The results that are depicted are preliminary results but they show that this approach is very promising since it clearly demonstrates that our graph representation is able to deal with the variability of pedestrian pose.

Kernel basis pursuit

Estimating a non-uniformly sampled function from a set of learning points is a classical regression problem. Kernel methods have been widely used in this context, but every problem leads to two major tasks: optimizing the kernel and setting the fitness-regularization compromise. This article presents a new method to estimate a function from noisy learning points in the context of RKHS (Reproducing Kernel Hilbert Space). We introduce the Kernel Basis Pursuit algorithm, which enables us to build a L 1 -regularized-multiple-kernel estimator. The general idea is to decompose the function to learn on a sparse-optimal set of spanning functions. Our implementation relies on the Least Absolute Shrinkage and Selection Operator (LASSO) formulation and on the Least Angle Regression Stepwise (LARS) solver. The computation ofthe full regularization path, through the LARS, will enable us to propose new adaptive criteria to find an optimal fitness-regularization compromise. Finally, we aim at proposing a fast parameter-free method to estimate non-uniform-sampled functions.

Translation-invariant classification of non-stationary signals

Non-stationary signal classification is a complex problem. This problem becomes even more difficult if we add the following hypothesis: each signal includes a discriminant waveform, the time location of which is random and unknown. This is a problem that may arise in Brain Computer Interfaces (BCI) or in electroencephalogram recordings of patients prone to epilepsy. The aim of this article is to provide a new graph-based representation for classifying this kind of signals. This representation characterizes the waveform without reference to the absolute time location of the pattern in the signal. We will show that it is possible to create such a signal description using graphs on a time-scale or time-frequency signal representation. The definition of an inner product between graphs is then required to implement kernel methods algorithms like Support Vector Machines. Our experimental results shows that this approach is very promising and performs very well on real-world datasets.