Using Riemannian geometry for SSVEP-based Brain Computer Interface
Extracted source text · arXiv 1501.03227
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% arXiv
% to convert from Elsevier to arXiv
% change {Figures/ -> {
% change figure} -> figure*}
% change figure size -> 0.8
% change table} -> table*}
% change bibliography style
% change hyperref settings
% cd ../manuscript/Figures
% cp $(grep '^[[:space:]]*\\pgfimage' ../../arxiv/OnlineSSVEP-BCI-RiemannianGeometry.tex | cut -d'{' -f2 | cut -d'}' -f1) ../../arxiv/
% ls $(grep '^[[:space:]]*\\includegraphics' ../../arxiv/OnlineSSVEP-BCI-RiemannianGeometry.tex | cut -d'{' -f2 | cut -d'}' -f1) and copy the *-eps-converted-to.pdf
%change the includegraphics figure with their pgfimage equivalent.
\documentclass[11pt,a4paper]{article}
%% Elsevier style
%\documentclass[5p]{elsarticle} %,authoryear
%\documentclass[5p]{elsarticle} % once accepted
% \bibliographystyle{elsarticle-num}
%\journal{Preprint}
%% IEEE style
%\documentclass[letterpaper, 10 pt, conference]{ieeeconf} % Comment this line out if you need a4paper
%\documentclass[a4paper, 10pt, conference]{ieeeconf} % Use this line for a4 paper
%\bibliographystyle{IEEEtran}
%\IEEEoverridecommandlockouts % This command is only needed if
% you want to use the \thanks command
%\overrideIEEEmargins % Needed to meet printer requirements.
% See the \addtolength command later in the file to balance the column lengths
% on the last page of the document
% The following packages can be found on http:\\www.ctan.org
\usepackage{graphicx} % for pdf, bitmapped graphics files
\usepackage{pgf}
\usepackage{amssymb,amsmath,amsfonts}
\usepackage{bbm}
%\usepackage{epsfig} % for postscript graphics files
%\usepackage{mathptmx} % assumes new font selection scheme installed
%\usepackage{times} % assumes new font selection scheme installed
%\usepackage{amsmath} % assumes amsmath package installed
%\usepackage{amssymb} % assumes amsmath package installed
\usepackage{multirow}
\usepackage{booktabs}
%\usepackage{cite}
\usepackage{color}
\usepackage[font=footnotesize]{caption}
\usepackage[font=footnotesize]{subcaption}
\usepackage{algorithm}
\usepackage[noend]{algpseudocode}
\usepackage{hyperref}
\usepackage{xcolor}
\definecolor{lnkcol}{rgb}{0,0,0.93}
%\captionsetup[table]{textfont=it,position=top}
%\captionsetup[subtable]{textfont=sf}
%\usepackage{float}
%\usepackage{epstopdf}
%\floatstyle{boxed}
%\restylefloat{figure}
\hypersetup{
% bookmarks=true, % show bookmarks bar?
% unicode=true, % non-Latin characters in Acrobat?s bookmarks
% pdftoolbar=true, % show Acrobat?s toolbar?
% pdfmenubar=true, % show Acrobat?s menu?
% pdffitwindow=true, % page fit to window when opened
pdfauthor={Emmanuel Kalunga, Sylvain Chevallier and Quentin Barthelemy}, % author
pdfsubject={Using Riemannian geometry for SSVEP-based Brain Computer Interface},
pdfnewwindow=true, % links in new window
pdfkeywords={Riemannian geometry, Online, Asynchronous, Brain-Computer Interfaces, Steady State Visually Evoked Potentials},
colorlinks=true, % false: boxed links; true: colored links
linkcolor=lnkcol, % color of internal links
citecolor=lnkcol, % color of links to bibliography
% urlcolor=extcol % color of external links
}
% macros
%\newcommand{\x}{\ensuremath{\hat{x}}}
\newcommand{\ra}[1]{\renewcommand{\arraystretch}{#1}}
% macros
%\newcommand{\x}{\ensuremath{\hat{x}}}
\newcommand{\X}{\ensuremath{X}} % Trial signal matrix (training) : X
\newcommand{\onlineX}{\ensuremath{\chi}} % Trial signal matrix (training) : X
\newcommand{\Y}{\ensuremath{Y}} % Trial signal matrix (test): Y
\newcommand{\totd}{\ensuremath{{m}}} % dimension of the manifold : m
\newcommand{\chI}{\ensuremath{c}} %channel or channel index: c
\newcommand{\dc}{\ensuremath{C}} % dimension nb of electrode : C
\newcommand{\si}{\ensuremath{n}} %sample or sample index: n
\newcommand{\dt}{\ensuremath{N}} % dimension nb of samples (trial length): N
\newcommand{\dr}{\ensuremath{\mathcal{N}}} % recording length: N
\newcommand{\nb}{\ensuremath{i}} %index of a number : i
\newcommand{\Nb}{\ensuremath{I}} %a number of objects : I
\newcommand{\ci}{\ensuremath{k}} %class or class index: k
\newcommand{\dK}{\ensuremath{K}} %nb classes for SSVEP: K
\newcommand{\df}{\ensuremath{f}} %frequency SSVEP index: f
\newcommand{\dF}{\ensuremath{F}} %dimension nb of frequencies for SSVEP: F
\newcommand{\ti}{\ensuremath{i}} %trial or trial index: i
\newcommand{\dT}{\ensuremath{I}} %dimension nb of trials : I
\newcommand{\setindex}{\ensuremath{\mathcal{I}}} %Set of indeces of trials corresponding to a specific class I
\newcommand{\dd}{\ensuremath{D}} %dimension nb of segment(epochs) for classif: D
\newcommand{\tj}{\ensuremath{j}} %trial or trial index: j
\newcommand{\Dindex}{\ensuremath{\mathcal{J}}} %Set of indeces within the last \dd segments (epoch)
\newcommand{\fpnorm}{\ensuremath{p}} %fixed pont normalisation factor : p
\newcommand{\cue}{\ensuremath{\tau_0}} %Cue on set time: tau
\newcommand{\deltaN}{\ensuremath{\Delta \si}} % Sliding step size
\newcommand{\ws}{\ensuremath{w}} % Sliding window size w
\newcommand{\di}{\ensuremath{d}} % Sliding segment(epoch) index d
\newcommand{\Xs}[1] {\ensuremath{\X_{\operatorname{\text{#1}}}}} % Sliding EEG Segment(epoch)
\newcommand{\Xsig}{\ensuremath{\overrightarrow{\X}}}
\newcommand{\GenMa}{\ensuremath{\mathcal{M}}} %generic manifold M
\newcommand{\Ma}{\ensuremath{\mathcal{M}_{\dc}}} % Riemannian manifold of SPD matrices
\newcommand{\Sy}{\ensuremath{\mathcal{S}_{\dc}}} % Euclidean space of symmetric matrices
\renewcommand{\P}{\ensuremath{\Sigma}} % A point of the Riemannian manifold, was P
\renewcommand{\S}{\ensuremath{\Theta}} % A point of the tangent space, was S
\newcommand{\ev}{\ensuremath{\lambda}} % an eigenvector
\newcommand{\distR}{\ensuremath{\delta}} % Riemannian metric
\newcommand{\Rm}{\ensuremath{\mu}} % Riemannian mean
\newcommand{\EX}{\ensuremath{\omega}} % Expected vector of \X
\newcommand{\tarco}{\ensuremath{\Gamma}} % target covariance in shrinkage estimator
\newcommand{\mle}{\ensuremath{\hat{\ell}}} % maximum likelihood estimator of fixed point cov
\newcommand{\fmle}{\ensuremath{g}} % function of fixed point cov MLE
\newcommand{\supertrial}{\ensuremath{Z}} % super-trial signal of Barachant.
\newcommand{\pocc}{\ensuremath{\rho}} % occurence probability for classification
\newcommand{\pthres}{\ensuremath{\vartheta}} % probability threshold for classification
\newcommand{\clout}{\ensuremath{\ci^*}} % class output from classification algo 1 :f*
\newcommand{\cloutOn}{\ensuremath{\tilde{\ci}}} % class output from classification algo 2 :~f
\newcommand{\cloutmp}{\ensuremath{\bar{\ci}}} % class output from classification algo 2 :f**
\newcommand{\clarray}{\ensuremath{\mathcal{K}}} % class output from classification algo 2 :~f
% \newcommand{\mle}{\ensuremath{\hat{\mathbf{M}}}}
%\newcommand{\SCM}{\ensuremath{\mathbf{S}}}
% \newcommand{\Po}{\ensuremath{\mathcal{P}}} no need, use \Ma instead
%\newcommand{\Rm}{\ensuremath{\mathfrak{G}}} % I prefer to avoid German char
\newcommand{\x}{\ensuremath{x}} % I prefer non-bold notations for vectors
\newcommand{\xb}{\ensuremath{\bar{x}}}
% \newcommand{\x}{\ensuremath{\mathbf{x}}}
% \newcommand{\xb}{\ensuremath{\mathbf{\bar{x}}}}
\newcommand{\eye}{\ensuremath{\mathbbm{I}}}
\newcommand{\unitary}{\ensuremath{\mathbbm{1}}}
\newcommand{\argmin}{\arg\!\min}
%\newcommand{\argmax}{\arg\!\max}
\DeclareMathOperator*{\argmax}{arg\,max}
\renewcommand{\Re}{\ensuremath{\mathbb{R}}}
\newcommand{\Log}[1]{\ensuremath{\operatorname{Log}_{#1}}}
\newcommand{\Exp}[1]{\ensuremath{\operatorname{Exp}_{#1}}}
\newcommand{\Expm}{\ensuremath{\operatorname{Exp}}}
\newcommand{\Logm}{\ensuremath{\operatorname{Log}}}
\newcommand{\cov}[1]{\ensuremath{\hat{\Sigma}_{\operatorname{#1}}}}
\newcommand{\disvec}{\ensuremath{\widetilde{\delta}}} %\vec{\delta}}}
\newcommand{\diffdist}{\ensuremath{\delta}} % \pmb{\delta}}}
%\newcommand{\classM}[1]{\ensuremath{\P^{({#1})}_{\mathfrak{G}}}}
%\renewcommand{\cite}[1]{\citep{#1}}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% correct bad hyphenation here
\hyphenation{op-tical net-works semi-conduc-tor}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{document}
\begin{titlepage}
\newcommand{\HRule}{\rule{\linewidth}{0.5mm}} % Defines a new command for the horizontal lines, change thickness here
\center % Center everything on the page
%----------------------------------------------------------------------------------------
% HEADING SECTIONS
%----------------------------------------------------------------------------------------
\textsc{\LARGE Tshwane University of Technology}\\ % Name of your university/college
\textsc{\LARGE \&}\\ % Name of your university/college
\textsc{\LARGE Universit\'e de Versailles Saint-Quentin}\\ % Name of your university/college
\textsc{\LARGE \&}\\ % Name of your university/college
\textsc{\LARGE Mensia Technologies}\\[1.5cm] % Name of your university/college
% \textsc{\Large Major Heading}\\[0.5cm] % Major heading such as course name
% \textsc{\large Minor Heading}\\[0.5cm] % Minor heading such as course title
%----------------------------------------------------------------------------------------
% TITLE SECTION
%----------------------------------------------------------------------------------------
\HRule \\[0.4cm]
{ \huge \bfseries Research Report:\\Using Riemannian geometry for SSVEP-based Brain Computer Interface}\\[0.4cm] % Title of your document
\HRule \\[1.5cm]
%----------------------------------------------------------------------------------------
% AUTHOR SECTION
%----------------------------------------------------------------------------------------
\begin{minipage}{0.4\textwidth}
\begin{flushleft} \large
\emph{Authors:}\\
Emmanuel \textsc{Kalunga}\\[1.3cm] % Your name
Sylvain \textsc{Chevallier}\\[1.5cm]
Quentin \textsc{Barth\'elemy}\\
\end{flushleft}
\end{minipage}
~
\begin{minipage}{0.4\textwidth}
\begin{flushright} \small
\emph{Affiliations:} \\
Department of Electrical Engineering, F'SATIE, TUT, Pretoria, South Africa\\[0.5cm]
Laboratoire d'Ing\'{e}nierie des Syst\`{e}mes de Versailles, UVSQ, Velizy, France\\[0.5cm]
Mensia Technologies, ICM, H\^{o}pital de la Piti\'{e}-Salp\^{e}tri\`{e}re, Paris, France\\
\end{flushright}
\end{minipage}\\[1cm]
% If you don't want a supervisor, uncomment the two lines below and remove the section above
%\Large \emph{Author:}\\
%John \textsc{Smith}\\[3cm] % Your name
%----------------------------------------------------------------------------------------
% DATE SECTION
%----------------------------------------------------------------------------------------
{\large \today}\\[1cm] % Date, change the \today to a set date if you want to be precise
%----------------------------------------------------------------------------------------
% LOGO SECTION
%----------------------------------------------------------------------------------------
\pgfimage[interpolate=true,width=0.15\linewidth]{uvsqlogo} % Include a department/university logo - this
\hfill
\pgfimage[interpolate=true,width=0.2\linewidth]{logo-TUT} % Include a department/university logo - this
\hfill
\pgfimage[interpolate=true,width=0.2\linewidth]{logoMensia} % Include a department/university logo - this will require the graphicx package
%----------------------------------------------------------------------------------------
\vfill % Fill the rest of the page with whitespace
\end{titlepage}
%% arXiv
\title{Using Riemannian geometry for SSVEP-based Brain Computer Interface}
\author{
Emmanuel K. Kalunga\thanks{Department of Electrical Engineering, French South African Institute of Technology, Tshwane University of Technology, Pretoria 0001, South Africa. \texttt{emmanuelkalunga.k@gmail.com}} \and
Sylvain Chevallier\thanks{Laboratoire d'Ing\'{e}nierie des Syst\`{e}mes de Versailles, Universit\'{e} de Versailles Saint-Quentin, Velizy 78140, France. \texttt{sylvain.chevallier@uvsq.fr}} \and
Quentin Barth\'{e}lemy\thanks{Mensia Technologies, ICM, H\^{o}pital de la Piti\'{e}-Salp\^{e}tri\`{e}re, 75013 Paris, France. \texttt{quentin.barthelemy@mensiatech.com}}
}
\maketitle
\begin{abstract}
Riemannian geometry has been applied to Brain Computer Interface (BCI) for brain signals classification yielding promising results.
Studying electroencephalographic (EEG) signals from their associated covariance matrices allows a mitigation of common sources of variability (electronic, electrical, biological) by constructing a representation which is invariant to these perturbations. %with several invariances. %mitigate
While working in Euclidean space with covariance matrices is known to be error-prone, one might take advantage of algorithmic advances in information geometry and matrix manifold to implement methods for Symmetric Positive-Definite (SPD) matrices.
This paper proposes a comprehensive review of the actual tools of information geometry and how they could be applied on covariance matrices of EEG.
In practice, covariance matrices should be estimated, thus a thorough study of all estimators is conducted on real EEG dataset.
As a main contribution, this paper proposes an online implementation of a classifier in the Riemannian space and its subsequent assessment in Steady-State Visually Evoked Potential (SSVEP) experimentations.
\end{abstract}
{\small \textbf{Keywords:}\\
Riemannian geometry, Online, Asynchronous, Brain-Computer Interfaces, Steady State Visually Evoked Potentials.}
%% IEEE
%Emmanuel: On devra se mettre d'acord sur un titre. Entre temps je propose celui-ci
% \title{\LARGE \bf
% Online SSVEP-based BCI using Riemannian Geometry
% }
% %A Review of Riemannian Manifold of SPD Matrices Applied to BCI
% %A Review of Riemannian Geometry Applied to BCI
% \author{Emmanuel K. Kalunga$^{1,2}$, Sylvain Chevallier$^{1}$, Karim Djouani$^{2}$, Eric Monacelli$^{1}$, Yskandar Hamam$^{2}$% <-this % stops a space
% %\thanks{*This work was not supported by any organization}% <-this % stops a space
% \thanks{$^{1}$With Laboratoire d'Ing\'{e}nierie des Syst\`{e}mes de Versailles, Universit\'{e} de Versailles Saint-Quentin, Velizy 78140, France
% {\tt\small emmanuelkalunga.k@gmail.com}}%
% \thanks{$^{2}$With the Department of Electrical Engineering and the French South African Institute of Technology, Tshwane University of Technology,
% Pretoria 0001, South Africa}%
% }
% \begin{abstract}
% %insister sur le fait qu'on a besoin d'estimateur et d'algo qui prennent en compte la geometrie des SPD.
% % Insister sur le besoin de online pour la BCI
% Riemannian Geometry has been applied to BCI for brain signals classification yielding promising results.
% Studying electroencephalographic (EEG) signals from their associated covariance matrices allows to harness common sources of variability (electronical, electrical, biological) by constructing a representation with several invariances. %mitigate
% While working in Euclidean space with covariance matrices is known to be error-prone, one might take advantage of algorithmic advances in information geometry and matrix manifold to implement algorithms for Symetric Positive-Definite (SPD) matrices
% This paper proposes a comprehensive review of the actual tools of information geometry and how they could be applied on covariance matrices of EEG.
% A thorough study of all covariance estimators on experimental results is provided.
% We propose to an online implement of the classifier in the Riemannian space and the subsequent evaluation on Steady-State Visually Evoked Potential dataset.
% %This work present the general concept of Riemannian Geometry applied to BCI, it reviews the works where Riemannian Geometry has been used for brain signals classification, and demonstrates the efficiency of these methods on the online classification of Steady State Visual Evoked Potential.
% \end{abstract}
% \maketitle
% \thispagestyle{empty}
% \pagestyle{empty}
%% Elsevier
% \begin{frontmatter}
% \title{Using Riemannian geometry for SSVEP-based Brain Computer Interface}
% \author[tut,lisv]{Emmanuel K. Kalunga}
% \ead{emmanuelkalunga.k@gmail.com}
% \author[lisv]{Sylvain Chevallier\corref{cor}}
% \ead{sylvain.chevallier@uvsq.fr}
% \cortext[cor]{Corresponding author}
% \author[men]{Quentin Barth\'{e}lemy}
% \ead{quentin.barthelemy@mensiatech.com}
% \address[tut]{Department of Electrical Engineering and \\
% the French South African Institute of Technology, \\
% Tshwane University of Technology, \\
% Pretoria 0001, South Africa}
% \address[lisv]{Laboratoire d'Ing\'{e}nierie des Syst\`{e}mes de Versailles, \\
% Universit\'{e} de Versailles Saint-Quentin, \\
% Velizy 78140, France}
% \address[men]{Mensia Technologies,\\
% ICM, H\^{o}pital de la Piti\'{e}-Salp\^{e}tri\`{e}re,\\
% 75013 Paris, France}
% \begin{abstract}
% Riemannian geometry has been applied to Brain Computer Interface (BCI) for brain signals classification yielding promising results.
% Studying electroencephalographic (EEG) signals from their associated covariance matrices allows a mitigation of common sources of variability (electronic, electrical, biological) by constructing a representation which is invariant to these perturbations. %with several invariances. %mitigate
% While working in Euclidean space with covariance matrices is known to be error-prone, one might take advantage of algorithmic advances in information geometry and matrix manifold to implement methods for Symmetric Positive-Definite (SPD) matrices.
% This paper proposes a comprehensive review of the actual tools of information geometry and how they could be applied on covariance matrices of EEG.
% In practice, covariance matrices should be estimated, thus a thorough study of all estimators is conducted on real EEG dataset.
% As a main contribution, this paper proposes an online implementation of a classifier in the Riemannian space and its subsequent assessment in Steady-State Visually Evoked Potential (SSVEP) experimentations.
% \end{abstract}
% \begin{keyword}
% Riemannian geometry \sep Online /sep Asynchronous \sep Brain-Computer Interfaces \sep Steady State Visually Evoked Potentials
% \end{keyword}
% \end{frontmatter}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Introduction}
Human-machine interactions without relying on muscular capabilities is possible with Brain-Computer Interfaces (BCI)~\cite{wolpaw2002bcireview,PFU10,ZAN11}.
%Brain-Computer Interfaces (BCI) allow people to interact with a machine without requiring muscular capabilities~\cite{wolpaw2002bcireview}.
They are the focus of a large scientific interest~\cite{VID73,BAY00,LOP09,HUA11,JRA11,TU12}, especially those based on electro-encephalography (EEG) \cite{NIE04,Blankertz2006}.
From the large literature based on the BCI competition datasets~\cite{BLA04,BLA06,TAN12}, one can identified the two most challenging BCI problems: on one hand, the inter-individual variability plagues the models and lead to BCI-inefficiency effect~\cite{ALL10,VID10,HAM12}, on the other hand the intra-individual changes calls for the development of online algorithms and adaptive systems following the evolution of the subject's brain waves~\cite{Obermaier2001,SHE06,Lenhardt2008,Bin2009AnOnlineMultiCh,ZHU11}. %Guclcuturk2010
To alleviate these variations, several signal processing and machine learning techniques have been proposed, such as filtering, regularization or clustering~\cite{Lee2003,kalunga2013ssvep,Lu2010Regular} without marking out an obvious ``best candidate'' methodology. %identifying a clear mainstream methodology
A common vision is shared by all the most efficient approaches to reduce signal variabilities: they are applied on covariance matrices instead of working in the input signal space.
%One can note that to reduce the signal variabilities, a common, all the most efficient approaches share on a common vision: they rely on the covariance representations instead of processing directly the signal.
Common Spatial Pattern (CSP)~\cite{BLA08b,LOT11,ANG12,YAN12a}, which is the most known preprocessing technique in BCI, try to maximize the covariance of one class while minimizing the covariance of the other one.
Similarly, Principal Components Analysis (PCA)~\cite{BLA04,BLA06}, also applied for spatial filtering in BCI, is based on the estimation of covariance matrices.
Canonical Correlation Analysis (CCA)~\cite{Hardoon2004} is another example of a technique relying on covariance estimates successfully applied on EEG for spatial filtering~\cite{Lin2007,Bin2009AnOnlineMultiCh,kalunga2013ssvep}.
Given two sets of signals, CCA aims at finding the projection space that maximizes their cross-covariance while jointly minimizing their covariance.
%, while minimizing their intra covariance. % minimal
%Among the tools for reducing the signal variabilities, a common one is to rely on the covariance matrices instead of directly processing the signal.
%have been often used for the processing and classification of brain signals.
%In fact, covariance matrices have been employed in the most successful approaches to process and classify BCI signals.
%Common Spatial Pattern (CSP), the most common preprocessing technique in BCI, relies on the maximization or minimization of signals covariance in one class or another.
%Principal Components Analysis (PCA), also used in BCI for spatial filtering, uses covariance matrices in their computation.
%Canonical Correlation Analysis (CCA), is another technique that relies on covariance matrices and successfully used in EEG filtering and classification.
%It find a projection space that maximises the covariance of two set of data, while keeping their intra covariance minimal \cite{Hardoon2004,kalunga2013ssvep}.
Covariance matrices are also found in classifiers such as the Linear Discriminant Analysis (LDA), which is largely used in BCI.
In all cases, they are handled as elements of an Euclidean space.
However, being Symmetric and Positive-Definite (SPD), covariance matrices lie on a subset of the Euclidean space, with reduced dimensionality and specific properties, the \emph{Riemannian manifold}. % $\Ma^{m}$.
Considering covariance matrices in their original space would reduce the search area for an optimization problem \cite{absil2009optimization}. % from the entire $\Re^{n}$ space to a subspace $\Ma^{m}$ with $m < n$.
As Riemannian manifolds inherently define a metric, the distance between SPD matrices takes into account the space where they lie on; approximating it to an Euclidean space introduce inaccuracies and result in ill-conditioned matrices.
Recently, studies have been done to consider covariance matrices obtained from multichannel brain signals in their original space \cite{barachant2010riemannian, barachant2012multiclass}.
Covariance matrices are the input features of the BCI system and the classifier algorithms rely on Riemannian metric for partitioning the feature space.
The authors propose to build specific covariance matrices in order to emphasize the spatial and temporal information of the multichannel brain signals.
%Covariance matrices are not only used for spatial filtering or calculation purposes as part of a more complex algorithm;
%they are used on their own as classification features, providing both spatial and frequential information about the multichannel brain signal, and in a space that allows an accurate evaluation of these features, resulting in a simplified algorithm.
The outcome of this approach is a simple processing toolchain, which achieve state-of-the-art classification performances.
This paper introduces an online version of the \emph{minimum distance to Riemannian mean} algorithm~\cite{barachant2012multiclass}, with an application on Steady-State Visually Evoked Potential (SSVEP) recordings.
In SSVEP, the subjects concentrate on stimuli blinking with fixed frequencies; depending on the focus of their attention, brain waves will arise with the same phase and frequency than the stimulus chosen by the subject.
%In this paper, an online implementation of the algorithm is proposed and its efficiency is demonstrated.
The proposed online implementation improves the performances in the BCI task, as the parameters of the classifier adapt to new, unlabeled samples.
Such adaptation takes into account the qualitative changes in the brain responses and the acquired signals, caused respectively by internal and external factors, e.g. user fatigue and slight variations in experimental settings.
The proposed online implementation is similar to the unsupervised or semi-unsupervised learning scheme proposed in~\cite{panicker2010adaptation,schettini2014self}; that has the potential of shortening (or even removing) the calibration phase.
The proposed algorithm applies a similar approach to dynamic stopping criterion used in \cite{verschore2012dynamic} to increase the speed of the BCI system.
This approach allows to dynamically determine the trial length and ensure robustness in classification results.
When working with covariance matrices, a crucial point is to correctly estimate the covariance when the number of samples is small or heavily corrupted by noise.
Several approaches have been proposed to build the covariance matrices, relying on normalization or regularization of the sample covariances.
To assess the quality of the covariance matrices obtained from EEG samples, a comparative study of these estimators is conducted.
Hence, the contributions of this works are:
\begin{itemize}
\item a comprehensive review of the literature on Riemannian geometry applied to EEG and time-series,
\item an online and asynchronous classification algorithm for SSVEP-based BCI,
\item a thorough analysis of the covariance estimators and their impact on tools derived from information geometry.
\end{itemize}
The paper is divided as follows:
Section~\ref{sec:stateoftheart} reviews application of Riemannian geometry to machine learning.
Section~\ref{sec:riem_geom} presents concepts of Riemannian geometry relevant to this work and estimators of covariance.
In Section~\ref{sec:proposed_algo}, the proposed classification algorithm for online SSVEP is introduced and the experimental results are presented in Section~\ref{sec:exp_valid}.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{State of the Art}
\label{sec:stateoftheart}
%zz: a reorganiser un peu. mettre le point de vue des different algo en avant, il faut trouver un axe directeur.
% Eviter les phrases qui commencent par le \cite{} ou In \cite{}. Il y en a trop
% Une proposition de plan:
% - La geom?trie riemannienne pour am?liorer les algorithmes des machine learning existant:
% - les algorithmes supervis?s : CSP (barachant2010common, samek2013robust, samek2014information) & SVM (barachant2012bci, yger2013review, jayasumana2013kernel )
% - les algorithmes de r?duction de dimensions : LE (goh2008clustering, ...), LLE, H-LLE, dictionary learning (xie2013nonlinear), channel selection (barachant2011channel)
% - les outils qui travaillent directement dans l'espace des vari?t?s riemaniennes :
% - MDRM, barachant2012multiclass, barachant2010riemannian congedo2013new
% - riemannian potato barachant2013riemannian
% - weighted distance li2009eeg li2012electroencephalogram
Information geometry provides useful tools for various machine learning and optimization problems.
In machine learning, SPD matrices % and probability distribution functions (pdf)
have been used in various applications where features and data are only considered in the Euclidean space. % were explored with tools from Riemannian geometry.
Indeed, covariance matrices lie in the space of SPD matrices which is a subset of the Euclidean space when considered with the scalar product.
But the same space of SPD matrices, endowed with a differential structure, induces a Riemannian manifold.
%Probability distribution functions of observed data and their covariance matrices, which are SPD, are elements of a Riemannian manifold. %zz: j'ai chang? le paragraphe car il parlait de riemannian structure, je ne suis pas sur de la d?finition d'une structure riemannienne. Est ce que tu pourrais m'indiquer o? tu as lu ?a ?
% *La geom?trie riemannienne pour am?liorer les algorithmes des machine learning existant
%To overcome the inadequacy of applying algorithms and operations defined in Euclidian space on data lying on Riemannian manifolds, %zz:idem pour les structure
%Use of kernels to use algo developed for Euclidean space
Riemannian geometry can improve machine learning algorithms, taking explicitly into consideration the underlying structure of the considered space.
%Riemannian geometry have been used to improve machine learning algorithms, most of which were mainly developed for data in euclidean space, to apply them on data lying on Riemannian manifold.
Three kind of approaches in the literature uses the data geometry in machine learning.
The first one relies on the mapping of the Riemannian manifold onto an Euclidean vector space. % where Euclidean geometry applies.
One such mapping, called logarithmic mapping, exist between the manifold and its tangent space, which is an Euclidean space, and has been used in classification task for BCI~\cite{barachant2012bci,BAR13}.
%The Riemannian Log map induces a direct mapping on the tangent space, those tangent spaces are Euclidean.
%Mapping on the tangent space using the Riemannian Log map \cite{barachant2012bci} could be used as tangent spaces on Riemannian manifold are known to be Euclidean.
Other kernels have been applied successfully to this end: Stein kernel, Log-Euclidean kernels as well as their normalized versions~\cite{yger2013review}.
The main idea is to map the input data to a high dimensional feature space, providing a rich and hopefully linearly separable representation.
The so-called kernel trick is to provide a kernel function, which computes an inner product in the feature space directly from points lying in the input space, defining a Reproducing Kernel Hilbert Space (RKHS).
The family of kernels defined on the Riemannian manifold allows implementing extension of all kernel-based methods, such as SVM, kernel-PCA or kernel $k$-means~\cite{jayasumana2013kernel}.
%Using kernels that are symmetric positive definite and adapted from Euclidean metrics to the Riemannian geodesic metrics to map SPD matrices to higher dimension linear Hilbert space works well on kernel-based algorithms such as SVM and kernel PCA applied to data on Riemannian manifold
Apart from the kernel approaches, once the data are mapped onto a vector space, any machine learning algorithm working in Euclidean space, such as LDA, could be applied~\cite{barachant2012multiclass}.
% Reformulation and adaptation of algorithmes supervis?s :
A second kind of machine learning approach exploit the underlying geometry of the data.
Instead of mapping the data to an Euclidean space, either a tangent space or a RKHS, the algorithms are adapted to Riemannian space.
% Using kernels makes the numerous existing machine learning algorithms readily usable for data originally on Riemannian manifold. However, some features might be squeezed out during the mapping process.
% Another avenue has been explored, where the data are left in their original Riemannian space, whereas the algorithms initially developed for Euclidean space are adapted to Riemannian space.
For instance, sparse coding algorithm have been adapted to Riemannian manifold, using the geodesic distance to estimate the data point and its sparse estimate~\cite{xie2013nonlinear}.
%When data points and atoms used to generate them belong to a Riemannian space $\Ma$, they do no support vector space structure. Hence data points are not estimated as linear combination of atoms (as it is the case in sparse coding), rather with a nonlinear sparse coding; and the error (distance) between the data point and its estimate from sparse coding is a geodesic \cite{xie2013nonlinear}.
Similarly nonlinear dimensionality reduction techniques have been adapted to Riemannian manifold, such as Laplacian Eigenmaps (LE), Locally Linear Embedding (LLE), and Hessian LLE.
This adaptation was used to cluster data using their pdfs \cite{goh2008unsupervised} or covariance matrices \cite{goh2008clustering} as features.
Another example is the adaptation of interpolation and filtering of data to Riemannian space performed in \cite{PEN06}, where an affine-invariant Riemannian metric is also proposed to offer a geodesically complete manifold i.e a manifold with no edge and no singular point that can be reached in finite time.
%The authors of these techniques assume that the data features belonging to $\Ma$ are distributed in submanifolds of $\Ma$; and hence each of the submanifold will be mapped to a different point in $\Re^n$. % zz: the notation
%\textit{k}-disconnected union of \textit{m} \textit{k}-connected submanifolds of $\Ma$; and hence each of the submanifold will be mapped to a different point in $\Re^m$. % zz: the notation
% * les outils qui travaillent directement dans l'espace des vari?t?s riemaniennes : (Newly developed algo for features riemanian manifold)
In the last kind of approach, instead of adapting existing algorithm from Euclidean to Riemannian geometry, new algorithms are developed directly for Riemannian manifolds.
% MDRM
The \emph{minimum distance to Riemannian mean} (MDRM) relies on a Riemannian metric to implement a multi-class classifier and have been applied on EEG.
%an equivalent of $k$-mean in % between covariance matrices of EEG to classify trials from multiclass motor imagery BCI.
New EEG trials are assigned to the class whose average covariance matrix is the closest to the trial covariance matrix~\cite{barachant2012multiclass}.
The MDRM classification can be preceded by a filtering of covariance matrices, like in~\cite{barachant2010riemannian} where covariance matrices are filtered with LDA component in the tangent space, then brought back to the Riemannian space for classification with MDRM.
Another example is the \emph{Riemannian Potato} \cite{barachant2013riemannian}, an unsupervised and adaptive artifact detection method, providing an online adaptive EEG filtering (i.e outliers removal).
Incoming signals are rejected if their covariance matrix lies beyond a predefined distance z-score from the mean covariance matrix, computed from a sliding window.
% which is defined adaptively along signal recording.
With the same objective of achieving robustness to noise that affect covariance matrices, Riemannian geometry is used to solve divergence functions of pdfs~\cite{amari2010information}.
This allows to reformulate the CSP as the maximization of the divergence between the distributions of data from two different classes corresponding to two cognitive states~\cite{samek2013robust, samek2014information}.
Using the \emph{beta divergence} the obtained CSP is robust to outliers in sample covariance matrices and this algorithm is successfully applied to EEG filtering for BCI.
Riemannian metrics are also used for EEG channel selection~\cite{barachant2011channel} and the selection of the most discriminatory spatial filters in CSP~\cite{barachant2010common}.
%%Feature (filters) selection in CSP
%In \cite{barachant2010common}, Riemannian distance is used to select the most discriminatory spatial filters in CSP from eigenvectors of the sum of class mean covariance matrices -- which are obtained with a Riemannian mean.
%It is established that each eigenvalue corresponding to each filter (eigenvector) contribute to the Riemannian distance between the class mean covariance matrices.
%Filters whose eigenvalue contribute the most are selected.
%%channel selection
%In \cite{barachant2011channel}, the authors propose a channel selection technique using backward elimination principle.
%Channels that maximize the Riemannian distance between Riemannian mean covariance matrices of different classes are kept.
%Riemannian Geometry for Evoked potential and ERP
Applications of Riemannian geometry to BCI mentioned thus far are focusing on motor imagery (MI) paradigm.
In MI experiment, the subject is asked to imagine a movement (usually hand, feet or tongue), generating Event-Related Synchronization and Desynchronization (ERD/ERS) in pre-motor brain area.
Riemannian BCI is well suited for MI experiment as the spatial information linked with synchronization is directly embedded in covariance matrices obtained from multichannel recordings.
%needed to identify regions of synchronization and regions of desynchronization is well embedded in covariance matrices obtained from the multichannel observed data.
However, for BCI that rely on Evoked Potential such as SSVEP or Event Related Potential (ERP), as P300, both frequential and temporal information are needed; the spatial covariance matrix does not contain these information.
%-By an extended definition od SCM
To apply Riemannian geometry to SSVEP and ERP, the sample covariance matrices can be defined from a rearrangement of the recorded data.
The rearrangement is done such that the temporal or frequency information are captured~\cite{congedo2013new}.
%-By new definition of riemannian distance
With similar motivations, \cite{li2009eeg} and \cite{li2012electroencephalogram} defined a new Riemannian distance between SPD matrices that would take into account a weighting factor on matrices.
They use this new distance as a dissimilarity between weighted matrices of power spectral density to classify EEG into different sleep state by $k$-nearest neighbors.
%Matrices of power spectral density are function of the frequency band $\omega$, and within each EEG epoch varying $\omega$ results in a curve on the manifold. the proposed metric measures the distance between such curves.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Covariance Matrices and their Geometry}
\label{sec:riem_geom}
This section presents some formal definitions for the information geometry concepts used in this paper.
The link with the covariance matrices is explicated in Section~\ref{sec:cov_mat}, along with the covariance estimators proposed in the literature.
\subsection{Riemannian Manifold}
\label{sec:notdef}
An $\totd$-dimensional \emph{manifold} $\GenMa$ is a Hausdorff space for which every point has a neighborhood that is homeomorphic to an open subset of $\Re^\totd$ \cite{lee2010introduction, jost2011riemannian}.
%\textcolor{red}{Remarque: je ne suis pas un expert en geometrie differentielle, mais m = C(C+1)/2. Donc ya un souci ici.}% $\Omega$
%Emmanuel :-) resoolu
When a tangent space is defined at each point, $\GenMa$ is called a differential manifold.
A \emph{geodesic} $\gamma$ is the shortest smooth curve between two points, $\P_1$ and $\P_2$.
The tangent space $T_\P\GenMa$ at point $\P$ is the vector space spanned by the tangent vectors of all geodesics on $\GenMa$ passing through $\P$. %On a Riemannian space,
%A \emph{geodesic} is the shortest smooth curve between two points, $P_1$ and $P_2$.
A \emph{Riemannian} manifold is a manifold endowed with an inner product defined on the tangent space, which varies smoothly from point to point. %at each point is a finite-dimensional Euclidean space.
For the rest of this paper, we will restrict to the analysis of the manifold $\Ma$ of the $\dc\times\dc$ symmetric positive definite matrices, defined as:
\begin{equation}
\label{eq:rm}
\Ma = \left\{ \P \in \Re^{\dc\times\dc} : \P = \P^\intercal \text{ and } x^\intercal \P x > 0, \forall x \in \Re^{\dc} \backslash 0 \right\} \ . \nonumber
\end{equation}
The tangent space $T_{\P}\Ma$ is identified to the Euclidean space of symmetric matrices:
\begin{equation}
\label{eq:sy}
\Sy = \left\{ \S \in \Re^{\dc\times\dc} : \S = \S^\intercal \right\} \ . \nonumber
\end{equation}
The dimension of the manifold $\Ma$, and its tangent space $T_{\P}\Ma$, is $\totd = \dc(\dc+1)/2$.
The mapping from a point $\S_i$ of the tangent space to the manifold is called the exponential mapping $\Exp{\P}(\S_i)$: $T_{\P}\Ma \rightarrow \Ma$ and is defined as:
\begin{equation}
\label{eq:exp_r}
\Exp{\P}(\S_i) = \P^{\frac{1}{2}}\Expm(\P^{-\frac{1}{2}}\S_i \P^{-\frac{1}{2}})\P^{\frac{1}{2}} \ .
\end{equation}
Its inverse mapping, from the manifold to the tangent space is the logarithmic mapping $\Log{\P}(\P_i)$: $\Ma \rightarrow T_{\P}\Ma$ and is defined as:
% The mapping from the Riemannian space to the tangent space is defined by the Riemannian Log map operator $\Log{\P}(\P_i)$: $\Ma \rightarrow T_{\P}\Ma$.
% Its inverse (back to the riemannian space) is the Riemannian Exp map operator $\Exp{\P}(\S_i)$: $T_{\P}\Ma \rightarrow \Ma$.
%This isomorphism is defined by Eq.~ \eqref{eq:exp_r} and~\eqref{eq:log_r}, they are diffeomorphism
\begin{equation}
\label{eq:log_r}
\Log{\P}(\P_i) = \P^{\frac{1}{2}}\Logm(\P^{-\frac{1}{2}}\P_i \P^{-\frac{1}{2}})\P^{\frac{1}{2}} \ .
\end{equation}
$\Expm(\cdot)$ and $\Logm(\cdot)$ are the the matrix exponential and logarithm respectively.
The computation of these operators is straightforward for SPD matrices of $\Ma$. %for symmetric matrices of $\Sy$. % in $T_{\P}\Ma$ or $\Ma$.
They are obtained from their eigenvalue decomposition (EVD):
\begin{displaymath}
\begin{split}
\P &= U\, \text{diag}(\ev_1, \dots, \ev_\dc)\, U^\intercal \ , \\
\Expm(\P) &= U\, \text{diag}(\log(\ev_1), \dots, \log(\ev_\dc))\, U^\intercal \ , \\
\Logm(\P) &= U\, \text{diag}(\exp(\ev_1), \dots, \exp(\ev_\dc))\, U^\intercal \ ,
\end{split}
\end{displaymath}
where $\ev_1, \dots, \ev_\dc$ are the eigenvalues and $U$ the matrix of eigenvectors of $\P$.
As any SPD matrix can be diagonalized with strictly positive eigenvalues, $\Logm(\cdot)$ is always defined.
Similarly the square root $\P^{\frac{1}{2}}$ is obtained as:
\begin{equation}
\label{eq:sqm}
\P^{\frac{1}{2}} = U\, \text{diag}(\ev_1^{\frac{1}{2}}, \dots, \ev_\dc^{\frac{1}{2}})\, U^\intercal \ , \nonumber
\end{equation}
and is unique. The same goes for $\P^{-\frac{1}{2}}$.
The tangent vector of the geodesic $\gamma(t)$ between $\P_1$ and $\P_2$, where $\gamma(0) = \P_1$ and $\gamma(1) = \P_2$ is defined as $v = \overrightarrow{\P_1 \P_2} = \Log{\P_1}(\P_2)$.
A Riemannian distance between $\P_1$ and $\P_2$ can thus be defined as~\cite{moakher2005differential}:
\begin{equation}
\label{eq:dist_r}
\distR(\P_1,\P_2) = \|\Logm(\P_1^{-1}\P_2)\|_F = \left[\sum_{\chI=1}^{\dc}\log^{2}\lambda_{\chI}\right]^{1/2} ,
\end{equation}
where $\lambda_\chI$, $\chI = 1, \dots, \dc$, are the eigenvalues of $\P_1^{-1}\P_2$.
From Eq.~\eqref{eq:dist_r}, the geometric mean of $\Nb$ points $\P_\nb$ on the manifold, $\nb = 1, \dots , \Nb$, can be defined as the point that minimizes the sum of squared distances to all $\P_\nb$:
\begin{equation}
\label{mean_r}
\Rm(\P_1, \dots, \P_\Nb) = \argmin_{\P \in \Ma} \sum_{\nb=1}^{\Nb} \distR^2(\P_\nb,\P) \ .
\end{equation}
%\|\log(P_k^{-1}P)\|_F^2
This geometric mean has no closed form, and can be computed iteratively \cite{fletcher2004principal}.
\subsection{Covariance Matrix Estimation}
\label{sec:cov_mat}
%Let $X \in \Re^{p \times n}$ denotes a sample of random variables, where $p$ is the number of variables and $n$ the number of samples, $X = [\x_1 \x_2 \dots \x_n]$
Let $\x_\si \in \Re^{\dc}$, $\si=1, \ldots, \dt$, denotes a sample of a multichannel EEG trial recorded on $\dc$ electrodes.
$\dt$ is the trial length.
Let $\X \in \Re^{\dc\times\dt}$ be the EEG trial such as $\X = [\x_1, \ldots, \x_{\dt} ]$.
Under the hypothesis that all $\dt$ samples $\x_\si$ are randomly drawn from a distribution, it follows that $\X$ is a variable of random vectors and its expected vector is $\EX = E\{ \X \}$ \cite{fukunaga1990introduction}. %Fukunaga page 18
The covariance matrix of the random vector $\X$ is defined by $\P = E\{ (\X - \EX)(\X - \EX)^{\intercal} \}$ and is unknown, thus an estimate $\cov{ }$ should be computed.
% The covariance matrix $\P$ of EEG multichannel recording is usually unknown and an estimate $\cov{ }$ should be computed.
%There are many techniques that are used to an estimate $\cov{ }$.
The choice of the appropriate estimator is crucial to verify that the obtained covariance matrices fulfill the following properties: they should be accurate, SPD and well-conditioned.
The last property requires that the ratio between the maximum and minimum singular value is not too large.
Moreover, to ensure the computational stability of the algorithm, the estimator should provide full-rank matrices, and its inversion should not amplify estimation errors.
\subsubsection{Sample Covariance Matrix Estimator}
The most usual estimator is the empirical \emph{sample covariance matrix} (SCM), defined as:
\begin{equation}
\begin{split}
\cov{scm} & = \frac{1}{\dt-1} \sum_{\si=1}^{\dt} (\x_\si - \xb)(\x_\si - \xb)^\intercal
\\
& = \frac{1}{\dt-1} \X\left( \eye_\dt -\frac{1}{\dt}\unitary_{\dt} \unitary_{\dt}^\intercal \right) \X^\intercal \ ,
\end{split}
\end{equation}
where $\xb$ is the sample mean vector $\xb = \frac{1}{\dt} \sum_{n=1}^\dt \x_\si $.
In the matrix notation, $\eye_{\dt}$ is the $\dt\times\dt$ identity matrix and $\unitary_{\dt}$ is the vector $[1, \ldots, 1]$.
The SCM is often normalized as:
\begin{equation}
\label{eq:nscme}
\cov{nscm} = \frac{\dc}{\dt} \sum_{\si=1}^{\dt} \frac{(\x_\si - \xb)(\x_\si - \xb)^\intercal }{\sigma_{\x_\si}^2} \ ,
\end{equation}
$\sigma_{\x_\si}^2$ is the inter-channel variance at time $\si$. Other normalization techniques could be used.
This estimation is fast and computationally simple.
However when $\dc \approx \dt$, the SCM is not a good estimator of the true covariance.
In the case $\dc > \dt$, the SCM is not even full rank.
\subsubsection{Shrinkage Covariance Matrix Estimators}
To overcome these shortcomings, the shrinkage estimators have been developed as a weighted combination of the SCM and a target covariance matrix, which is often chosen to be close to the identity matrix, \textit{i.e.} resulting from almost independent variables of unit variance.
\begin{equation}
\label{eq:shrink}
\cov{shrink} = \kappa \tarco +(1-\kappa)\cov{scm} \ ,
\end{equation}
where $0 \leqslant \kappa < 1$.
This estimator provides a regularized covariance that outperforms the empirical $\cov{scm}$ for small sample size, that is $\dc \approx \dt$.
The shrinkage estimator has the same eigenvectors as the SCM, but the extreme eigenvalues are modified \textit{i.e.} the estimator is shrunk or elongated toward the average.
The different shrinkage estimators differ in their definition of the target covariance matrix $\tarco$.
Ledoit and Wolf~\cite{ledoit2004well} have proposed $\tarco=\mathit{v}\eye_{\dc}$, with $\mathit{v} = \text{Tr}(\cov{scm} \eye_{\dc})$.
Blankertz~\cite{blankertz2011single} defines $\tarco$ also as $\mathit{v}\eye_{\dc}$ but with $\mathit{v} = \frac{\text{Tr}(\cov{scm})}{\totd}$, $\totd$ being the dimension of the covariance matrices space.
Sch\"{a}fer proposes several ways of defining $\tarco$ depending on the observed $\cov{scm}$~\cite{schafer2005shrinkage}.
%The shrinkage estimator shrinks (or enlarges) the sample covariance matrix toward a more \textcolor{red}{circular covariance}.
%When $\lambda = 1$, the estimated covariance matrix $\hat{\Sigma}$ does not take into consideration the empirical data; and when $\lambda=0$, $\hat{\Sigma}$ is not regularized.
%
%\subsection{Ledoit Shrinkage}
%Ledoit and Wolf developed a shrinkage estimator $\hat{\Sigma}$ of the covariance $\Sigma$ as a linear combination of the identity matrix and the sample covariance matrix, $\hat{\Sigma}=\rho_1I+\rho_2\SCM $ that minimizes the expected quadratic loss $E[\| \hat{\Sigma}-\Sigma \|]$. $\rho_1$ and $\rho_2$ are the shrinkage parameters defined by scalar functions of the covariance matrix $\Sigma$. Since estimator $\Sigma$ is unobserved, these parameters are estimated from the empirical covariance matrix $\SCM$ as described in [Ledoit 2004].
%\subsection{Sch\"{a}fer Shrinkage}
%Sch\"{a}fer and Strimmer proposed a shrinkage estimator based on Ledoit and Wolf work [Ledoit2003]. Their estimator is a weighted combination of the sample covariance matrix $\SCM$ and an estimated covariance target $\mathbf{T}$: $\hat{\Sigma}=\lambda \mathbf{T}+(1-\lambda)\SCM$. $\lambda$ is taken to be the argument that minimizes the squared error loss function, i.e. the Frobenius norm: $L(\lambda) = \| \hat{\Sigma} - \Sigma \|_F^2$. This estimator offers a systematic way to obtain a regularized $\hat{\Sigma}$ that outperforms the empirical $SCM$ and the theoretical $\mathbf{T}$, based on the knowledge of data.
%\subsection{Blankertz Shrinkage}
\subsubsection{Fixed Point Covariance Matrix Estimator}
The Fixed Point Covariance Matrix~\cite{pascal123theoretical} is based on the maximum likelihood estimator $\mle$ which is a solution to the following equation:
\begin{equation}
\label{eq:max_lik}
\cov{fp} = \mle = \frac{\dc}{\dt} \sum_{\si=1}^\dt \left( \frac{(\x_\si - \xb)(\x_\si - \xb)^\intercal }{(\x_\si - \xb)^\intercal \mle^{-1} (\x_\si - \xb)} \right) \ .
\end{equation}
%where $p$ is the normalization: $p = tr(\cov{scm})$.
%This equation admits a single fixed point $\hat{\mathbf{M}}_{fp}$ as solution.
As there is no closed form expression to Eq.~\eqref{eq:max_lik}, it can be written as a function of $\mle$: $\fmle(\mle) = \cov{fp}$.
%\begin{equation}
%\label{eq:max_lik_func}
%\fmle(\mle) = \frac{p}{\dc} \sum_{i=1}^\dt \left( \frac{(\x_i-\xb)(\x_i-\xb)^\intercal }{(\x_i-\xb)^\intercal \mle^{-1} (\x_i-\xb)} \right) \ .
%\end{equation}
$\fmle$ admits a single \emph{fixed point} $\mle^*$, where $\fmle(\mle^*) = \mle^*$, which is a solution to Eq.~\eqref{eq:max_lik}.
Using $\mle_{0}:=\cov{nscm}$ as the initial value of $\mle $, it is solved recursively as $\mle_{t} \underset{t \to \infty}{\longrightarrow} \mle^{*}$.
%*Fixed Point Covariance matrix estimation
%*Blankertz Shrinkage covariance
%
%It is about the quality of Covariance matrices and their SPD properties + non singular, full rank
%
%Many economic problems require a covariance matrix estimator that is not only invertible, but also well-conditioned (that is, inverting it does not amplify estimation error). For large-dimensional covariance matrices, the usual estimator - the sample covariance matrix - is typically not well-conditioned and may not even be invertible. This paper introduces an estimator that is both well-conditioned and more accurate than the sample covariance matrix asymptotically. This estimator is distribution-free and has a simple explicit formula that is easy to compute and interpret. It is the asymptotically optimal convex combination of the sample covariance matrix with the identity matrix. Optimality is meant with respect to a quadratic loss function, asymptotically as the number of observations and the number of variables go to infinity together. Extensive Monte-Carlo confirm that the asymptotic results tend to hold well in finite sample. [Ledoit]
%*Sample Covariance Matrix
%*Normalized Covariance Matrix (Proposed by Barachant)
%*Shrinkage estimation (Ledoit)
%*Fixed Point Covariance matrix estimation
%*Blankertz Shrinkage covariance
%*J. Schaefer Shrinkage
%\section{Existing works}
%\label{sec:review}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Online Adaptation\\of the Riemannian Classifier}
\label{sec:proposed_algo}
Concerning Riemannian classification of SSVEP, the offline methodology has been explained in \cite{congedo2013new}.
In this paper, we propose an online classifier for SSVEP, composed of an offline training step and an online and asynchronous test step.
This analysis is performed for each subject independently.
\subsection{Offline Riemannian Classification}
The proposed classifier relies on the Minimum Distance to Riemannian Mean (MDRM) introduced in~\cite{barachant2010riemannian} and extended in~\cite{congedo2013new} for possible applications on SSVEP signals.
Let consider an experimental SSVEP setup with $\dF$ stimulus blinking at $\dF$ different frequencies. It is a multiclass classification with $\dK=\dF+1$ classes: one class per stimulus and one resting state class.
The covariance matrices are estimated from a modified version of the input signal $\X$: %as explained in~\cite{congedo2013new}
%Let $X \in \Re^{p \times n}$ be the recorded data with $p$ the number of channels, and $n$ the number of sample per trial.
%The extended data are defined as:
\begin{equation}
\X \in \Re^{\dc \times \dt} \rightarrow
\begin{bmatrix}
X_{\text{freq}_1}\\
\vdots \\
X_{\text{freq}_{\dF}} \\
\end{bmatrix}
\in \Re^{\dF\dc \times \dt} \ ,
\label{eq:ext_data}
\end{equation}
where $X_{\text{freq}_\df}$ is the input signal $\X$ band-pass filtered around frequency $\text{freq}_\df$, $\df=1, \ldots, \dF$.
Thus the resulting covariance matrix $\cov{}$ belongs to $\Re^{\dF\dc \times \dF\dc}$. Henceforth, all EEG signals will be considered as filtered and modified by Eq.~\eqref{eq:ext_data}.
From $\dT$ labelled training trials $ \left\{ \X_{\ti} \right\}_{\ti=1}^{\dT}$ recorded per subject, $\dK$ centers of classes $\P_{\Rm}^{(\ci)}$ are estimated using Algorithm~\ref{alg:r_mean}.
%The MDRM is then used for classification, as shown in Algorithm~\ref{alg:mdrm}.
When an unlabelled test trial $\Y$ is given, it is classified as belonging to the class whose center $\P_{\Rm}^{(\ci)}$ is the closest to the trial's covariance matrix (Algorithm~\ref{alg:mdrm}, step \ref{op:decision}).
\begin{algorithm}
\caption{Offline Estimation of Riemannian Centers of Classes}
\label{alg:r_mean}
Inputs: $\X_{\ti} \in \Re^{\dF \dc\times\dt}$, for $\ti = 1, \ldots, \dT$, a set of labelled trials. \\
Inputs: $\setindex(\ci)$, a set of indices of trials belonging to class $\ci$. \\
Output: $\P_{\Rm}^{(\ci)}$, $\ci = 1, \ldots, \dK$, centers of classes.
\begin{algorithmic}[1]
\State Compute covariance matrices $\cov{\textit{\ti}}$ of $\X_\ti$ %, Eq.~\eqref{eq:shrink}
\label{op:cov_i}
\State \textbf{for} $\ci$ = 1 \textbf{to} $\dK$ \textbf{do}
\State \quad $\P_{\Rm}^{(\ci)}=\Rm(\cov{\textit{\ti}}:\ti \in \setindex(\ci))$ , Eq.~\eqref{mean_r}
\label{op:class_center}
\State \textbf{end}
\State \textbf{return} $\P_{\Rm}^{(\ci)}$
\end{algorithmic}
\end{algorithm}
%
\begin{algorithm}
\caption{Minimum Distance to Riemannian Mean}
\label{alg:mdrm}
Inputs: $\P_{\Rm}^{(\ci)}$, $\dK$ centers of classes from Algorithm~\ref{alg:r_mean}. \\
Input: $\Y \in \Re^{\dF \dc\times\dt}$, an unlabelled test trial. \\
Output: $\clout$, the predicted label of $\Y$.
\begin{algorithmic}[1]
\State Compute covariance matrix $\cov{}$ of $\Y$ %, Eq.~\eqref{eq:shrink}
\label{op:cov}
\State $\clout = \argmin_{\ci} \distR(\cov{}, \P_{\Rm}^{(\ci)})$
\label{op:decision}
\State \textbf{return} $\clout$
\end{algorithmic}
\end{algorithm}
\subsection{Curve-based Online Classification} %pas Gradient-based
\begin{algorithm}
\caption{Curve-based Online Classification}
\label{alg:online}
Inputs: hyper-parameters $\ws$, $\deltaN$, $\dd$, and $\pthres$.\\
Inputs: $\P_{\Rm}^{(\ci)}$, $\ci = 1, \ldots, \dK$, centers of classes from Algorithm~\ref{alg:r_mean} (offline training). \\
Inputs: Online EEG recording $\onlineX(\si)$. \\
Output: $\cloutOn(\si)$, online predicted class.
\begin{algorithmic}[1]
\State $\di=1$
\State \textbf{for} $\si = \ws$ \textbf{to} $\dr$ \textbf{step} $\deltaN$
\State \quad Epoch $\Xs{\di}$, Eq.~(\ref{eq:online_epoch}), and classify it with Algorithm~\ref{alg:mdrm}
\label{op3:epoch_and_classify}
\State \quad \textbf{if} $\di \geq \dd$ %$\si \geq \ws+(\dd-1)\deltaN$
\label{op3:test_len}
\State \quad \quad Find the most recurrent class in $\clarray = \clout_{\tj \in \Dindex(d)}$:
$\cloutmp = \argmax_{\ci} \pocc(\ci)$ , Eq.~\eqref{eq:occ_prob}
\label{op3:get_rho}
\State \quad \quad \textbf{if} $\pocc(\cloutmp)>\pthres$
\label{op3:test_rho}
\State \quad \quad \quad Compute $\disvec_{\cloutmp}$ , Eq.~\eqref{eq:dist_vec} %This line can be left out
\label{op3:get_disvec}
\State \quad \quad \quad \textbf{if} $\disvec_{\cloutmp} < 0$
\label{op3:test_disvec}
\State \quad \quad \quad \quad \textbf{return} $\cloutOn = \cloutmp$
\label{op3:valid_class}
\State \quad \quad \quad \textbf{end}
\State \quad \quad \textbf{end}
\State \quad \textbf{end}
\State \quad $\di = \di+1$
\label{op3:increment_d}
\State \textbf{end}
\end{algorithmic}
\end{algorithm}
%Sylvain: Est ce qu'on peut donner une formule pour calculer \pocc qui soit pas trop ridicule ?
%Emmanuel: Comme \pocc est une probabilite, sa formule serait (tu peux uncomment pour voir ce que ca donne:
%\[
%\pocc(j) = \frac{\textrm{number of outcome }j}{\textrm{total number of outcomes}}
%\]
%qui pourait s'ecrire aussi:
%\[
%\pocc(j) = \frac{1}{D}| \{ i \in \{ 1, \dots, \dd \}:\clout_i = j \} |
%\]
%ou peut etre tout simplement:
%\[
%\pocc(j) = \frac{1}{D} \sum_{i=1}^{\dd} \left[\clout_i = j \right]
%\]
%Sylvain: changer les parametres et dire qu'on prend les centers de classe.
%Emmanuel: C'est fait!
In offline synchronous BCI paradigm, cue onset are used as reference for the localization of a brain response, e.g. for an evoked potential.
Nonetheless most of the BCI applications are online and asynchronous; cue onsets are not known, thus designing online version of BCI algorithms, even well documented ones, is not a trivial task.
%In online and asynchronous experiment, there is no such cue onset and the adaptation of the algorithm for online processing is not trivial and often difficult.
The approach introduced here identifies a period (\textit{i.e.} interval) in the online EEG $\onlineX \in \Re^{\dF\dc\times\dr}$, where $\dr$ is the number of recorded samples, associated with a high probability (above threshold) of observing an SSVEP at a specific frequency, as illustrated in Algorithm~\ref{alg:online}. % (i.e. beyond a predefined threshold).
To locate this interval, we focus on the last $\dd$ recorded EEG overlapping epochs
$ \left\{ \Xs{\ensuremath{i}} \in \Re^{\dF\dc \times \ws} \right\}_{\tj \in \Dindex(d)} $ ,
with the set of indices $\Dindex(d) = \di-\dd +1,\ldots,\di-1,\di$;
where $\di$ is the index of the current epoch $\Xs{\di}$ in the online recording $\onlineX(n)$.
Epochs have size $\ws$, and the interval between two consecutive epochs is $\deltaN$, with $\ws > \deltaN $:
%Epochs $\Xs{\di}$ are obtained by sliding a window of size $\ws$ by a step of size $\deltaN$ every time $\deltaN$ new samples are available (step~\ref{op3:epoch}):
%\[ \Xs{\di} = \onlineX(\time) \Big|_{\time=\si-\ws,\ldots,\si} \]
\begin{equation}
\label{eq:online_epoch}
\Xs{\di} = \onlineX(\si-\ws,\ldots,\si) \ .
\end{equation}
%Consecutive $\Xs{\di}$ overlap, with $\ws > \deltaN $.
%To sum up, $\Xs{\di}$ slides over all the signal to give successive epochs for the online framework.
To obtain the first $\dd$ epochs $\Xs{\ensuremath{\tj \in \Dindex(d)}}$, at least $\ws+(\dd-1)\,\deltaN$ samples of $\onlineX$ should be recorded (step~\ref{op3:test_len}).
The classification outputs $\clout_{\tj \in \Dindex(d)}$ obtained by applying Algorithm~\ref{alg:mdrm} (step~\ref{op3:epoch_and_classify})
on $\Xs{\ensuremath{\tj \in \Dindex(d)}}$ are stored in a vector $\clarray$.
The class that occurs the most in $\clarray$ (step~\ref{op3:get_rho}), with an occurrence probability $\pocc(\ci)$ above a defined threshold $\pthres$
%(where $\pocc$ is a vector of size $\dF$ containing the probabilities of occurrence of each class in $\clarray$)
, is considered to be the class, denoted $\cloutmp$, of the ongoing EEG recording $\onlineX(\si)$. Vector $\pocc$ is defined as:
\begin{equation}
\label{eq:occ_prob}
\begin{split}
\pocc(\ci) = \frac{\#\{ \clout_{\tj \in \Dindex(d)} = \ci \}}{\dd} \ & , \text{ for } \ci = 1, \ldots, \dK,
\\
\text{and} \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \cloutmp = {\argmax}_{\ci} & \ \pocc(\ci) \ .
\end{split}
\end{equation}
If $\pthres$ is not reached within the last $\dd$ epochs, the classification output is held back, and the sliding process continues until $\pthres$ is reached.
%and with a new epoch $\Xs{\di+1}$ classified, and $\clarray$ updated with $\clout_{\di+1}$.
%The process carries on until $\pthres$ is reached.
In the last $\dd$ epochs, once a class $\cloutmp$ has been identified, a curve direction criterion is introduced to enforce the robustness of the result.
For class $\cloutmp$ to be validated, this criterion requires that the direction taken by the displacement of covariance matrices $\cov{\textit{\tj} \in \Dindex(\textit{\di})}$ be toward the center of class $\P_{\Rm}^{(\cloutmp)}$. %\classM{l}
Hence $\disvec_{\cloutmp}$, the sum of gradients (\textit{i.e.} differentials) of the curve made by distances from $\cov{ \textit{\tj} \in \Dindex(\textit{\di})}$ to $\P_{\Rm}^{(\cloutmp)}$ should be negative (step~\ref{op3:test_disvec}):
%%A window of size $w_s$ seconds is slided from $t = t_0 - w_s$ to $t_0$ with a step of $\Delta t$ seconds.
%%The most represented class in the $\dd$ last classifications, with an occurrence probability $\pocc(j)$ reaching a predetermined threshold $\pthres$, is considered to be the class of the current EEG samples.
%%%Sylvain : inclure une ref a l'algo avec la ligne
%%If $\pthres$ is not reached within $\dd$ epochs, no classification is done and the sliding window process goes on.
%%When considering $\dd$ epochs, if a specific class is selected several times in a row, a curve direction criterion is introduced to enforce the stability of the results. %validate the output. %gradient
%Once the classification result is more or less \emph{steady} on a specific class, a direction criterion is introduced to validate the output. %gradient
%Sylvain : idem, ref a la ligne dans l'algo
\begin{equation}
\label{eq:dist_vec}
\begin{split}
\disvec_{\cloutmp} &= \sum_{\tj \in \Dindex(\di)} \frac{\Delta \diffdist_{\cloutmp}(\tj)}{\Delta \tj}
\ = \sum_{\tj=\di-\dd+2}^{\di} \diffdist_{\cloutmp}(\tj) - \diffdist_{\cloutmp}(\tj-1)
\ < 0
\\
& \text{with} \ \ \ \ \diffdist_{\cloutmp}(\tj) = \frac{\distR (\cov{\textit{\tj}}, \P_{\Rm}^{(\cloutmp)} )}{\sum_{\ci=1}^{\dK} \distR (\cov{\textit{\tj}},\P_{\Rm}^{(\ci)})} \ .
\end{split}
\end{equation}
%$L$ is the number of classes, and $d_R(\cdot,\cdot)$ is the Riemannian distance between two covariance matrices at $\Delta_t = n$.\\
The occurrence criterion is inspired by the dynamic stopping of \cite{verschore2012dynamic}; there is no fixed trial length for classification.
The occurrence criterion ensures that the detected user intention is unaffected by any short time disturbances due to noise or subject's inattention, as presented in Algorithm~\ref{alg:online}.
This approach offers a good compromise to obtain robust results within a short and flexible time.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Experimental Validation}
\label{sec:exp_valid}
Covariance matrix estimators, Algorithms~\ref{alg:mdrm} and \ref{alg:online} are applied to SSVEP signals for offline and online analysis. This section presents the analysis and results obtained. %It is divided into 4 parts: the description of data used, the analysis and conclusion on the quality of covariance estimators, the effect of outliers, and lastly the analysis and results obtained with the proposed online algorithm.
\subsection{Data Description}
The signals are recorded from 12 subjects during an SSVEP experiment.
EEG are measured on $\dc=$ 8 channels: $O_Z$, $O_1$, $O_2$, $PO_Z$, $PO_3$, $PO_4$, $PO_7$, and $PO_8$.
The ground and the reference electrodes were placed respectively on $F_Z$ and the right hear mastoid.
The acquisition rate is $T_s = 256$ Hz on a gTec MobiLab amp.
The subjects are presented with $\dF=$ 3 visual target stimuli blinking respectively at $\text{freq}=$ 13Hz, 17Hz and 21Hz.
It is a $\dK=$ 4 classes BCI setup made of the $\dF=$ 3 stimulus classes and one resting class (no-SSVEP).
In a session, which lasts 5 minutes, 32 trials are recorded: 8 for each visual stimulus and 8 for the resting class.
The number of sessions recorded per subject varies from 2 to 5.
Thus the longest EEG recorded for a single subject is 25 minutes or 160 trials.
The trial length is 6 seconds, that is $\dt=6 \times T_s = 1536$ samples.
Since data are rearranged as detailed in \eqref{eq:ext_data}, trials $\X \in \Re^{\dF\dc \times \dt}$, where $\dF\dc = 24$ corresponding to 8 channels times 3 stimulus frequencies.
For each subject, a test set is made of 32 trials while the remaining trials (which might vary from 32 to 128) make up for the training set.
%Thus the resulting covariance matrix $\hat{\Sigma}$ belongs to $\Re^{24 \times 24}$.
\subsection{Covariance Estimators Comparison}
\begin{figure*}[t!]
\centering
\begin{minipage}[b]{0.8\linewidth}
\pgfimage[width=1\textwidth]{accuracy_errorbar-eps-converted-to.pdf}
% \includegraphics[width=1\textwidth]{accuracy_errorbar.eps}
\subcaption{}
\label{fig:acc_errorbar}
\end{minipage}
\begin{minipage}[b]{0.8\linewidth}
\pgfimage[width=1\textwidth]{itr-eps-converted-to.pdf}
% \includegraphics[width=1\textwidth]{itr.eps}
\subcaption{}
\label{fig:itr_errorbar}
\end{minipage}
\caption{Comparison of covariance estimators in terms of classification accuracy and Information Transfer Rate (ITR) obtained with MDRM with increasing EEG trial length. For each trial length, the average accuracy and ITR across all subjects and across all replications is shown. Bars indicate the standard deviation. (a) Accuracy in percentage, (b) ITR in bits/min.}
\label{fig:class_size}
\end{figure*}
%To perform classification on the Riemannian manifold, it is crucial to ensure that the covariance matrices obtained from the measured signal are symmetric positive-definite, that they are well conditioned and representative of the samples.
In this section, the effectiveness of covariance matrix estimators is evaluated for SSVEP signals.
The evaluation is done in terms of classification accuracy, information transfer rate (ITR), and integrated discrimination improvement (IDI), obtained by each estimator (see Section~\ref{sec:cov_mat}) while using the offline MDRM classifier.
%\textcolor{red}{QB:Si c est MDRM, toutes les comparaisons sont effectu\'ees sur algo offline, et non online. ce qui est incoherent avec le fait de decouper des epoch (on parle de epoch size)... Idem legende Fig 1!}.
The different conditioning of covariance matrices are also investigated.
%The estimators analyzed are described in.
%\subsubsection{Methods}
A bootstrapping with 1000 replications is performed to assess the performances of each estimator.
%From the pool of recorded samples available for each subject, a set of training data is randomly sampled with replacement; similarly, a set of test data is selected, with no overlap between the two sets. Class-mean matrices are obtained from the training set, and the classification performed on the test set. This is done for each matrix estimator (used to compute covariance matrix of samples), and the performance of each estimator are calculated. And the process is repeated 1000 times.
Estimators are compared on 10 trial lengths $t \in \{ 0.5, 1.0, \dots 5.0\}$ seconds, as these are known to affect the estimators performance. Here $\dt \in \{ 128, 256, \dots, 1280 \}$ is computed as $\dt=t \times T_s$.
Figure~\ref{fig:acc_errorbar} shows the classification accuracies of each estimator.
The increase in the accuracy can be attributed to the fact that the relevant patterns in EEG accumulate with the trial length, producing better estimation of the covariance matrices.
%It is as well due to the fact that an increase in sample size results in a
This is known to be particularly true for the SCM estimator and it could be seen in Figure~\ref{fig:acc_errorbar}.
It appears that shrinkage estimators (especially Ledoit and Sch\"afer) are less affected by the reduction of epoch sizes than the other estimators. This is a direct consequence of the regularization between the sample covariance matrices and the targeted (expected) covariance matrix of independent variables.
The information of Figure~\ref{fig:acc_errorbar} is congruent with Figure~\ref{fig:itr_errorbar}; ITR increases as the trial length decreases. The transfer rate is higher for shrinkage estimators with shorter epoch sizes, while maintaining correct accuracy.
For computational purposes, it is important to look at the matrix conditioning.
Figure~\ref{fig:eigenvalue_range} shows the ratio $\mathcal{C}$ between the largest and smallest eigenvalues: in well-conditioned matrices, $\mathcal{C}$ is small.
Shrinkage estimators offer better conditioned matrices while the SCM, NSCM, and Fixed Point matrices are ill-conditioned below 2 seconds of trial length, and may result in singular matrices.
\begin{figure*}[ht!]
\centering
\begin{minipage}[b]{0.8\linewidth}
\pgfimage[width=1\textwidth]{eigenvalue_range-eps-converted-to.pdf}
% \includegraphics[width=1\textwidth]{eigenvalue_range.eps}
\subcaption{}
\label{fig:eigenvalue_range}
\end{minipage}
\begin{minipage}[b]{0.8\linewidth}
\pgfimage[width=1\textwidth]{idi-eps-converted-to.pdf}
% \includegraphics[width=1\textwidth]{idi.eps}
\subcaption{}
\label{fig:idi}
\end{minipage}
\caption{(a) Covariance matrices condition expressed as the ratio $\mathcal{C}$ between largest and smallest eigenvalues for the different covariance estimators. The comparison is done for increasing EEG trial length. (b) Integrated discrimination improvement brought to the classification task by various estimators along varying trail length. The indicated IDI values are multiplied by $10^{2}$. $\cov{scm}$ is used as baseline. }
\label{fig:class_idi}
\end{figure*}
On Figure~\ref{fig:idi}, the Integrated Discrimination Improvement (IDI), as defined in \cite{pencina2008evaluating}, is computed for the different estimators and trial lengths.
The SCM is used as a reference for improvement, as this is the most popular estimator in the literature.
Negative IDI means a deterioration in the method discrimination ability.
It is clear that shrinkage estimators increase the discrimination power of the classifier.
However, despite being more complex than the SCM, the NSCM and the Fixed Point estimators decrease the discrimination ability of classifiers.
From Figures~\ref{fig:acc_errorbar} and~\ref{fig:idi}, it is apparent that the difference in performance between the SCM and shrinkage estimators reduces as the trial length increases.
The simplicity of the SCM plays a favorable role: it is an attractive method for longer trials.
However for trial lengths compatible with a reasonable ITR for man-machine interaction ($< 5$ sec), the shrinkage estimators offer better performances, with the Schafer estimator being the best.
The $p$-values under the hypothesis that there is no improvement (\textit{i.e.} IDI = 0) from one estimator to another are all inferior to $10^{-47}$, hence the improvement is significant.
It should be noted that the estimation of covariance matrices is a trade-off between the quality of the estimate and the computation time required; this should be considered for real-time processing.
\subsection{Effect of Outliers on Centers Estimation}
Outliers can affect the offline training of the $\dK$ centers of class $\P_{\Rm}^{(\ci)}$ by Algorithm~\ref{alg:r_mean}.
Figure~\ref{fig:tan_plan_s16&17} shows representations of training covariance matrices $\P_{\ti}$ in the tangent space ($\S_{\ti}$), projected at the mean of all training trials (black star).
To obtain this visualization, the first two principal components of a PCA applied on $\left\{ \S_{\ti} \right\}_{\ti=1}^{\dT}$ are selected.
In Figures~\ref{fig:tan_plan_s16_b} and~\ref{fig:tan_plan_s17_b}, a Riemannian potato~\cite{barachant2013riemannian} is applied with the Riemannian mean of all training matrices as a fixed reference.
Riemannian potato only removes highly corrupted signals, allowing to remove outlier trials that negatively impact the estimation of class centers.
%and thus this technique can help to successfully remove outliers trials, \textit{i.e.} trials corrupted by noise.
%all trials that were misclassified due to the noise are rejected . %extreme noise %that make the matrices go beyond the potato.
Applying this filter in an offline evaluation improves estimation of class centers for the subject with the lowest BCI performance.
Figure~\ref{fig:tan_plan_s16_b} illustrates how center estimates are affected by the filtering. Bad estimates of a class center will result in poor clustering. %17.
%\textcolor{red}{Ici, tu parles souvent de "classification accuracy" ou "classification results", mais a aucun moment tu ne donnes le pourcentage d augmentation de l accuracy que donne ce rejet d outliers (ce qui pourrait etre interessant comme resultat). Parles plutot de "clustering efficiency".}
%Emmanuel: J'ai plutot change ca en 'center estimation'
It should be noted that the Riemannian potato may have a negative effect if the different classes are well separated.
The Riemannian potato could reject trials that are close to the average of their class but that are far from the mean of all trials.
Thus, ``legitimate'' trials could be discarded, even if they have a positive impact on the classification results.
%It might also put in jeopardy a class that stands far away from the general mean by rejecting some of its trials that are well inside the potato of its own class-mean.
Avoiding this central bias is a possible improvement of the Riemannian potato algorithm. %A better filter should avoid this central bias.
\begin{figure*}[ht!]
\centering
\begin{minipage}[b]{0.49\linewidth}
\pgfimage[width=1\textwidth]{tan_plan_2D_s16_b-eps-converted-to.pdf}
%\includegraphics[width=1\textwidth]{tan_plan_2D_s16_b.eps}
\subcaption{}
\label{fig:tan_plan_s16_a}
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\pgfimage[width=1\textwidth]{tan_plan_2D_s16_a-eps-converted-to.pdf}
%\includegraphics[width=1\textwidth]{tan_plan_2D_s16_a.eps}
\subcaption{}
\label{fig:tan_plan_s16_b}
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\pgfimage[width=1\textwidth]{tan_plan_2D_s17_b-eps-converted-to.pdf}
%\includegraphics[width=1\textwidth]{tan_plan_2D_s17_b.eps}
\subcaption{}
\label{fig:tan_plan_s17_a}
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\pgfimage[width=1\textwidth]{tan_plan_2D_s17_a-eps-converted-to.pdf}
%\includegraphics[width=1\textwidth]{tan_plan_2D_s17_a.eps}
\subcaption{}
\label{fig:tan_plan_s17_b}
\end{minipage}
\caption{Scatter plot of covariance matrices for all trials mapped on the tangent space. The distance between each trial covariance matrix $\P_{\ti}$ and its Riemannian mean class $\P_{\Rm}^{(\ci)}$ is shown as connection line. The black star represents the Riemannian mean of all trials. % matrices regardless of their class.
Subject with lowest BCI performance, (a) before and (b) after Riemannian potato filtering.
Subject with highest BCI performance, (c) before and (d) after Riemannian potato filtering.}
\label{fig:tan_plan_s16&17}
% Sylvain: Est ce que c'est possible d'agrandir encore la patate Riemmanienne pour eviter qu'il reste encore un outlier dans la classe 17Hz pour le sujet 16 (lowest BCI perfs, sur la Figure b) ?
%Emmanuel: Oui, mais la patate ne sera plus comme definit pas Barachant (a 2.5 x deviation standard). Elle est maintenant a 1.12 x deviation standard
%Sylvain: Je pensais que les donnes serait plus etalee ainsi sur l'axe x, est possible de changer les valeur de debut et de fin du x-axis pour que les points des clusters soient repartis sur l'ensemble du plot ?
%Emmanuel: Oui, c'est possible. Je l'avais fais ainsi pour que l'image garde le meme aspect. C'est fait! J'ai du aussi changer ma PCA. Elle influe beaucoup sur la disposition des points.
%Sylvain: c'est normal. Les figures que tu as faites sont magnifiques !
\end{figure*}
\subsection{Classification Results}
In this section the classification performances are presented. It is divided into two parts: first, the relevance of identifying the latency between cue onset and SSVEP response is shown through the evaluation of the impact of SSVEP delays considered for offline classification. Then, performances of offline and online algorithms are compared and analysed.
\subsubsection{SSVEP Response Time Delay for Offline}
\label{sec:ssvep_response_delay}
Figure~\ref{fig:sync_delay} shows time signals of one trial from each class, measured at $Oz$, for the subject with highest BCI performance and for the subject with lowest BCI performance.
The figure displays the raw signal filtered at the 3 different stimulus frequency, each with a different color: 13Hz in blue, 17Hz in red, 21Hz in green.
%The signal is replicated 3 times with each replicate being the raw signal filtered at a different stimulus frequency. Each replicate is shown in a different color.
It is visible that synchronization for current trial frequency appears only 2 seconds after cue onset $\cue = 0 s$ on the $x$-axis of the plots.
In fact, before $\cue + 2$ seconds the signal is still synchronized with the previous trial frequencies.
In the signals from the subject with the highest BCI performance, synchronization at the trial stimulus frequency is predominant and permanent after $\cue + 2$, which eases the classification task.
Moreover, for the resting trial (no SSVEP), the power in all stimuli frequencies is sensibly decreased.
However, the poor predictive capability for the subject with lowest BCI performance could be explained by the fact that the predominance of the trial stimulus frequency is not always evident and not permanent.
Also, during the resting trial, the power in stimuli frequencies is not significantly decreased.
\begin{figure*}[ht!]
\centering
\begin{minipage}[b]{0.9\linewidth}
\pgfimage[width=1\textwidth]{ts17_f-eps-converted-to.pdf}
% \includegraphics[width=1\textwidth]{ts17_f.eps}
\subcaption{Example of trials for the subject with the highest BCI performance}
\end{minipage}
\begin{minipage}[b]{0.9\linewidth}
\pgfimage[width=1\textwidth]{ts16_f-eps-converted-to.pdf}
% \includegraphics[width=1\textwidth]{ts16_f.eps}
\subcaption{Example of trials for the subject with the lowest BCI performance}
\end{minipage}
\caption{Signal amplitude at each stimulus frequency, showing synchronization of EEG with respect to time (seconds). The raw signal of the trial measured on Oz is band filtered using a Butterworth of order 8 at each stimulus frequency and the resulting signals are shown in blue (dark grey) , green (grey), and red (light grey) for the same signal filtered respectively at 13, 17, and 21 Hz. The cue onset $\cue$ at time $0$ on the x-axis is shown with a vertical discontinued line. 4 trials are shown, one for each class. Signals from the subjects, (a) with highest BCI performance and (b) with lowest BCI performance.}
\label{fig:sync_delay}
\end{figure*}
Similar observation is made from Figure~\ref{fig:sync_delay_psd} where the time-frequency representations of EEG trials per class are plotted for the subjects with the highest and lowest BCI performances. % 11 (a) and 12 (b).
The time-frequency representation is obtained with a STFT on rectangular windowing.
Each plot has 4 rows, one for each class, displaying the power spectral density (averaged over all trials) in the 13Hz, 17Hz, and 21Hz frequency bands as a function of the trial time.
\begin{figure*}[ht!]
\centering
\begin{minipage}[b]{0.8\linewidth}
\pgfimage[width=1\textwidth]{mPSD_s17-eps-converted-to.pdf}
%\includegraphics[width=1\textwidth]{mPSD_s17.eps}
\subcaption{Time-frequency diagrams for the subject with the highest BCI performance}
\end{minipage}
\begin{minipage}[b]{0.8\linewidth}
\pgfimage[width=1\textwidth]{mPSD_s16-eps-converted-to.pdf}
%\includegraphics[width=1\textwidth]{mPSD_s16.eps}
\subcaption{Time-frequency diagrams for the subject with the lowest BCI performance}
\end{minipage}
\caption{Time-frequency representation of EEG from various classes showing synchronization at in frequency bands (indicated on y axis) on both side of the cue onset. The power of synchronization is shown in the color bar. Dark colors show strong synchronization. 4 trials (top to bottom) are represented, one for each class. Signals from the subjects, (a) with highest BCI performance and (b) with lowest BCI performance.}
\label{fig:sync_delay_psd}
\end{figure*}
An important increase in average classification accuracy (almost 10\%) could be obtained by taking the trial from 2 seconds after cue onset.
% Considering the delay between the cue onset and the synchronization in the appropriate frequency band by taking the EEG sample after 2 second after cue onset brings an increase in the average classification accuracy of up to 15\%.
It is therefore crucial to consider the latency between the trial's cue onset and the actual synchronization of SSVEP at stimulus frequency.
%locate the EEG signal window where classification results have the high confidence, while maintaining a decent response time in the system.
% without inserting a too long delay in the system response time.
%\textcolor{red}{Attention: reprendre la fin de cette section. On a l impression de faire du online optimis\'e. De plus, ce paragraphe ne devrait pas etre dans la section "Online Classification".}
\subsubsection{Online Classification Results}
%Code: SSVEP_Riemannian_multiclass.m: For synchronous results
% SSVEP_Riemannian_multiclass_cumulative.m: For asyncchronous
% SSVEP_Riemannian_multiclass_cumulative_curve.m: For asyncchronous (+ gradient curve) results
% SSVEP_Riemannian_multiclass.m: For offline opt.
% SSVEP_Riemannian_multiclass.m: For offline (state of the art).
% SSVEP_Riemannian_multiclass_naive.m: For raw application of MDRM on online and asych (naive)
In the offline synchronous analysis, the confident window for classification was found to be at least 2 seconds after the cue onset ($\cue+2$).
In an online asynchronous experiment, there is no cue onset, and the delay before SSVEP synchronization might defer from a trial to another and from a subject to another.
To locate the trust EEG region for the classification, $\dd$ and $\pthres$ are set respectively to 5 and 0.7 through cross validation. The epoch size is $\ws = 3.6 $ seconds, and the overlap is $\deltaN = 0.2$ second.
%A window of size $w_s$ seconds is slided from $t = t_0$, with a sliding step $\Delta_t$ seconds.
%The class that occurs the most in the $N$ latest classification (occurrence reaches a predetermined threshold $\rho$), is considered to be the class of the current EEG samples.
%If $\rho$ is not reached within $N$ slides, no classification is done and the sliding process continues.
% This approach is similar to the dynamic stopping of \cite{verschore2012dynamic}.
% In addition to the class occurrence criterion, the identified class $i$ is only validated if the displacement of covariance matrix $C_n$ at step $n$, is in the direction of the center of class $i$.
% Hence $\vec{\delta_i}$, the distance vector to class $i$ should be negative:
% \begin{equation}
% \vec{\delta_i} = \frac{\Delta \mathbf{x}_i(n)}{\Delta n} < 0
% \label{eq:dist_vec_online}
% \end{equation}
% \begin{equation}
% \mathbf{x}_i(n) = \frac{d_n(C_n,\varphi_i)}{\sum_j^{M}d_n(C_n,\varphi_j)}
% \label{eq:direction_online}
% \end{equation}
% $M$ is the number of classes, and $d_n(\cdot,\cdot)$ is the Riemannian distance between two covariance matrices at $\Delta_t = n$.\\
\begin{figure*}[ht!]
\centering
\begin{minipage}[b]{0.8\linewidth}
%\includegraphics[width=1\textwidth]{class_path1_s17v3.eps}
\pgfimage[width=1\textwidth]{class_path1_s17v3-eps-converted-to.pdf}
\subcaption{ }
\label{fig:class_path1_s17}
\end{minipage}
\begin{minipage}[b]{0.8\linewidth}
\pgfimage[width=1\textwidth]{class_path2_s17-eps-converted-to.pdf}
%\includegraphics[width=1\textwidth]{class_path2_s17.eps}
\subcaption{ }
\label{fig:class_path2_s17}
\end{minipage}
\caption{Covariance matrices trajectory during a 4-class SSVEP online recording. The circles represent class centers. The triangles mark the beginning in the experiment of a new trial whose class is indicated by the triangle's color. \ref{fig:class_path1_s17} shows the first 7 trials. The first 3 trials are from the resting class, the remaining are respectively class 13Hz, 17Hz, and 21Hz. \ref{fig:class_path2_s17} shows the entire recording. Data are from the subject with the highest BCI performance.}
\label{fig:class_path_s17}
\end{figure*}
%Equation~\eqref{eq:dist_vec} is derived from the observation of Figure~\ref{fig:class_path_s17}.
The observation of Figure~\ref{fig:class_path_s17} provides a visualization of the principle guiding the online implementation of Equation~\eqref{eq:dist_vec}.
This figure shows the trajectory on the tangent space taken by covariance matrices during a 4-class SSVEP experiment, and how they are classified epoch by epoch.
It can be seen (encircled in Figure~\ref{fig:class_path1_s17}) that a change in the SSVEP stimulus might not be detected instantaneously by the classifier.
The trials are erroneously attributed with confidence to the previous class.
The proposed online algorithm, described in Algorithm~\ref{alg:online}, mitigates this issue and increases the classification accuracy as shown in Table~\ref{tab:res}. %.and \eqref{eq:direction}
%Equation~\eqref{eq:dist_vec} alleviate this problem, and increase the classification accuracy as shown in Table~\ref{tab:res}. %and \eqref{eq:direction}
In the `Offline' column are the classification accuracies for each subject obtained by the application of the state-the-art MDRM as described in Algorithm~\ref{alg:mdrm}, which classifies full trials.
Column `Offline opt.' shows the results of this approach improved by taking into consideration the SSVEP response time delay as found in section~\ref{sec:ssvep_response_delay}: trials are taken 2 second after cue onset.
%This table shows that synchronous applications give better results, which is due to the fact that require extra information, \textit{i.e.} cue onsets. %but they are very limited in term of user interaction.
Columns `Online' and `Online + $\disvec_{\cloutmp}$' contain the performance of the online algorithm as described in Algorithm~\ref{alg:online}.
The `Online' column shows the results of the online algorithm without the curve direction criterion (\textit{i.e.}, without steps 6 to 11), and `Online + $\disvec_{\cloutmp}$' shows the improvement brought by this criterion.
The performances are in terms of average classification accuracy (acc(\%)) and average time taken into the trial before classification (delay(s)).
For the offline approaches, classification is only done at the end of the trial \textit{i.e.} 6 seconds after cue onset.
%User interaction are strongly limited by the synchronous paradigm, thus asynchronous algorithms are important for brain interfaces.
\\
The curve direction criterion increases the rejection of epochs that could be wrongly classified, it thus significantly increases the classification accuracy of the online algorithm (74.73\% to 79.13\%), while increasing the delay (1.006 s to 1.239 s) before classification.
When compared to the state-of-the-art MDRM, the curve-based classification yields better results both in terms of classification accuracy and delay before classification, \textit{i.e.} from 71.51\% to 79.13\% and from 6 s (or longer, depending on trial length $\dt$) to 1.239 s.
There is a small drop in classification accuracy from optimized offline (81.27\%) to online (79.13\%) algorithms.
However, classification outputs are reached faster with the online algorithm.
Moreover, the online algorithm can be applied in both synchronous and asynchronous paradigms, while the offline algorithms are limited to synchronous paradigms which provide strongly limited user interaction.
%Using $\vec{\delta_i}$ enhances the performance in asynchronous application (79.13\%) with shorter response time, even compared to synchronous results.
%\textcolor{red}{QB, Suggestion d ajout: These results validate the relevance and the efficiency of the proposed online approach, based on curve ... blablabla a completer.}
%E.K.: I will come back to this later.
\begin{table*}[ht!]
\centering
\ra{1}
\begin{tabular}{c|c|c||c|c|c|c|}
\cline{2-7}
& \multicolumn{1}{ c|}{Offline} &\multicolumn{1}{ c|| }{Offline opt.} & \multicolumn{2}{ c| }{Online} &\multicolumn{2}{ c| }{Online + $\disvec_{\cloutmp}$} \\ \cline{1-7}
\multicolumn{1}{ |c| }{Sub.} & acc(\%) & acc(\%) & acc(\%) & delay(s) & acc(\%) & delay(s)\\ \hline
\multicolumn{1}{ |c| }{1} & 62.50 & 73.44 & 70.31 & \textbf{1.081} & \textbf{71.37} & 1.338 \\ \hline
\multicolumn{1}{ |c| }{2} & 67.19 & 79.69 & \textbf{78.12} & \textbf{0.950} & 75.00 & 1.303 \\ \hline
\multicolumn{1}{ |c| }{3} & 85.94 & 85.94 & 87.50 & \textbf{0.950} & \textbf{89.06} & 0.972 \\ \hline
\multicolumn{1}{ |c| }{4} & 78.12 & 87.50 & 73.44 & \textbf{1.016} & \textbf{75.00} & 1.056 \\ \hline
\multicolumn{1}{ |c| }{5} & 70.31 & 68.75 & 70.31 & \textbf{1.013} & \textbf{70.31} & 1.319 \\ \hline
\multicolumn{1}{ |c| }{6} & 79.69 & 85.94 & \textbf{87.50} & \textbf{0.997} & 87.35 & 1.338 \\ \hline
\multicolumn{1}{ |c| }{7} & 78.12 & 88.54 & 84.38 & \textbf{1.004} & \textbf{85.42} & 1.144 \\ \hline
\multicolumn{1}{ |c| }{8} & 89.06 & 92.19 & 85.94 & \textbf{0.947} & \textbf{89.06} & 1.072 \\ \hline
\multicolumn{1}{ |c| }{9} & 62.50 & 70.31 & 67.19 & \textbf{0.922} & \textbf{75.00} & 1.244 \\ \hline
\multicolumn{1}{ |c| }{10} & 60.94 & 80.47 & 63.28 & \textbf{1.044} & \textbf{69.53} & 1.264 \\ \hline
\multicolumn{1}{ |c| }{11} & 43.75 & 65.62 & 59.38 & \textbf{1.109} & \textbf{68.75} & 1.375 \\ \hline
\multicolumn{1}{ |c| }{12} & 80.00 & 96.88 & 69.37 & \textbf{1.036} & \textbf{93.75} & 1.443 \\ \hline \hline
\multicolumn{1}{ |c| }{mean} & 71.51 & 81.27 & 74.73 & \textbf{1.006} & \textbf{79.13} & 1.239 \\ \hline
\end{tabular}
\caption{Average classification accuracies (acc(\%)) achieved using the state-of-the-art MDRM (Offline), the optimized version of the MDRM (Offline opt.), the online algorithm without the curve direction criterion (Online), and the complete online algorithm (Online + $\disvec_{\cloutmp}$). For the online algorithms, the delay between trial start and classification is also provided (delay(s)).}
%\caption{Average classification accuracy for each subject and time delay, from beginning of trial $t_0$, before valid classification output. The results obtained when using the cue onset $t_0$ as time reference and a fixed 2-seconds delay (synchronous) are compared to the results where no time reference is used (asynchronous), and enhanced by $\disvec_j$ (asynchronous + $\disvec_j$).}
\label{tab:res}
\end{table*}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Conclusion}
This work investigated the efficiency of Riemannian geometry when dealing with covariance matrices as classification features.
A novel algorithm, relying on the curve direction in the space of covariance EEG signals, was introduced and applied on a SSVEP classification task for a 4-class brain computer interface.
Different covariance matrix estimators were investigated and their robustness was assessed on multichannel SSVEP signals to ensure that the obtained matrices are accurate estimates of data covariance, are well conditioned, and verify the positive-definiteness property.
The Sch\"afer shrinkage estimator was found to be the best as it yielded the highest classification accuracy with the MDRM algorithm.
%when its obtained covariance matrices were used as features in MDRM classification.
The introduced online algorithm enhances the stability of the BCI system, balancing between the classification speed and the prediction accuracy.
%The algorithm provides a prediction as soon as an occurrence criterion is reached, thus adapting the trial length to variations during the experiment.
%Indeed, the classification results are validated as soon as, or only when the prescribed classification confidence, \textit{i.e.} the occurrence criterion, is reached, thus adapting the trial length to variations during the experiment.
Classification confidence over a number of epochs mitigates the short term perturbations in the experimental conditions and the attentional variations of the subject,
wile considering the curve direction overcomes the misclassification of EEG trials that are still synchronized with past stimuli frequencies at classification time.
%It allowed the detection misclassification of trials whose EEG signal are still synchronized on past stimuli frequencies at the time of classification.
The Riemannian geometric approach
and the determination of the Riemannian mean of classes,
offers a powerful tool to develop robust algorithms.
%Other than the trial length, it appeared important to have dynamic riemannian means of classes.
A possible extension to the proposed algorithm is to include a dynamic adaptation of the Riemannian mean for each class.
This dynamic adaptation was implemented, using the latest successfully classified covariance matrices to replace to oldest ones in the computation of the mean.
It is not reported in this paper as there was no significant improvement in the results, which could be attributed to the relatively short duration of the EEG recording sessions, which were no longer than 25 min.
The Riemannian means of classes do not change much over the sessions, but for longer recording time this adaptation could bring significant improvements.
% Though not included in this work, t
% with latest and successfully classified covariance matrices replacing the oldest in the computation of the means.
% However, the result was not improved.
% This could be attributed to the length of EEG recording sessions which were not longer than 25 min.
% The class riemannian means did not record significant change over the sessions.
% For longer sessions however, adaptive riemannian means could bring significant improvement.
% Our method benefits from the simplicity of Minimum Distance to Riemannian Means; there is few parameters.
% However few parameters, the EEG window size, the number of classifiers output considered for occurrence criterion, and the probability threshold should be predefined experimentally through cross validation.
% Many research avenues are still to be investigated, particularly to enhance the classification accuracy in subjects with low BCI performance.
%The perspectives of this work are mainly concerned with subjects with low BCI performance.
%Instead of using their own training data, which are hardly separable and negatively impacting the classification performance, we believe that a transfer learning approach could alleviate this issue.
%Models inherited from subjects with high BCI performance could help to improve the classifier accuracy, providing a reliable reference as starting point.
%Lastly, the Riemannian potato which was implemented to alleviate the impact of outliers showed some limitations, as it rejects trials whose covariance matrices stand beyond the Riemannian potato, yet close to their class mean.
%A filter without such central bias, taking into account the position of matrices depending of the class centers (\textit{i.e.} Riemannian mean), would improve to filter out the outliers.
%\section*{APPENDIX}
%Appendixes should appear before the acknowledgment.
\section*{Acknowledgment}
%
%
The authors would like to thank Louis~Mayaud from Mensia Technologies for his contribution in discussions that led to the completion of this work.
%\bibliographystyle{IEEEtran}
% {\small
\bibliography{BCIBib}
% }
\end{document}
\begin{thebibliography}{10}
\expandafter\ifx\csname url\endcsname\relax
\def\url#1{\texttt{#1}}\fi
\expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi
\expandafter\ifx\csname href\endcsname\relax
\def\href#1#2{#2} \def\path#1{#1}\fi
\bibitem{wolpaw2002bcireview}
J.~R. Wolpaw, N.~Birbaumer, D.~J. McFarland, G.~Pfurtscheller, T.~M. Vaughan,
{Brain-computer interfaces for communication and control}, Clinical
Neurophysiology 113~(6) (2002) 767--791.
\bibitem{PFU10}
G.~Pfurtscheller, B.~Z. Allison, C.~Brunner, G.~Bauernfeind,
T.~Solis-Escalante, R.~Scherer, T.~O. Zander, G.~Mueller-Putz, C.~Neuper,
N.~Birbaumer, The hybrid {BCI}, Frontiers in neuroscience 4.
\bibitem{ZAN11}
T.~O. Zander, C.~Kothe, Towards passive brain--computer interfaces: applying
brain--computer interface technology to human--machine systems in general,
Journal of Neural Engineering 8~(2) (2011) 025005+.
\bibitem{VID73}
J.~J. Vidal, Toward direct brain-computer communication., Annual review of
biophysics and bioengineering 2~(1) (1973) 157--180.
\bibitem{BAY00}
J.~D. Bayliss, D.~H. Ballard, Single trial p3 epoch recognition in a virtual
environment, Neurocomputing 32-33 (2000) 637--642.
\bibitem{LOP09}
M.~A. Lopez, H.~Pomares, F.~Pelayo, J.~Urquiza, J.~Perez, Evidences of
cognitive effects over auditory steady-state responses by means of artificial
neural networks and its use in brain--computer interfaces, Neurocomputing
72~(16-18) (2009) 3617--3623.
\bibitem{HUA11}
Y.~Huang, D.~Erdogmus, M.~Pavel, S.~Mathan, K.~E. Hild, A framework for rapid
visual image search using single-trial brain evoked responses, Neurocomputing
74~(12-13) (2011) 2041--2051.
\bibitem{JRA11}
N.~Jrad, M.~Congedo, R.~Phlypo, S.~Rousseau, R.~Flamary, F.~Yger,
A.~Rakotomamonjy, {sw-SVM}: sensor weighting support vector machines for
{EEG}-based brain--computer interfaces, Journal of Neural Engineering 8
(2011) 056004+.
\bibitem{TU12}
W.~Tu, S.~Sun, A subject transfer framework for {EEG} classification,
Neurocomputing 82 (2012) 109--116.
\bibitem{NIE04}
E.~Niedermeyer, F.~Lopes~da Silva, Electroencephalography: Basic Principles,
Clinical Applications, and Related Fields, 5th Edition, Lippincott Williams
\& Wilkins, 2004.
\bibitem{Blankertz2006}
B.~Blankertz, K.-R. Muller, D.~Krusienski, G.~Schalk, J.~Wolpaw, A.~Schlogl,
G.~Pfurtscheller, J.~Millan, M.~Schroder, N.~Birbaumer, {The BCI competition
III: validating alternative approaches to actual BCI problems}, Neural
Systems and Rehabilitation Engineering, IEEE Transactions on 14~(2) (2006)
153--159.
\bibitem{BLA04}
B.~Blankertz, K.-R.~R. M\"{u}ller, G.~Curio, T.~M. Vaughan, G.~Schalk, J.~R.
Wolpaw, A.~Schl\"{o}gl, C.~Neuper, G.~Pfurtscheller, T.~Hinterberger,
M.~Schr\"{o}der, N.~Birbaumer, The {BCI} competition 2003: progress and
perspectives in detection and discrimination of {EEG} single trials,
Biomedical Engineering, IEEE Transactions on 51~(6) (2004) 1044--1051.
\bibitem{BLA06}
B.~Blankertz, K.~R. Muller, D.~J. Krusienski, G.~Schalk, J.~R. Wolpaw,
A.~Schlogl, G.~Pfurtscheller, J.~Millan, M.~Schroder, N.~Birbaumer, The {BCI}
competition {III}: validating alternative approaches to actual {BCI}
problems, Neural Systems and Rehabilitation Engineering, IEEE Transactions on
14~(2) (2006) 153--159.
\bibitem{TAN12}
M.~Tangermann, K.-R. M{\"u}ller, A.~Aertsen, N.~Birbaumer, C.~Braun,
C.~Brunner, R.~Leeb, C.~Mehring, K.~J. Miller, G.~Mueller-Putz, G.~Nolte,
G.~Pfurtscheller, H.~Preissl, G.~Schalk, A.~Schl{\"o}gl, C.~Vidaurre,
S.~Waldert, B.~Blankertz, {Review of the BCI Competition IV}, Frontiers in
Neuroscience 6~(55).
\bibitem{ALL10}
B.~Z. Allison, C.~Neuper, Could anyone use a {BCI}?, in: D.~S. Tan, A.~Nijholt
(Eds.), Brain-Computer Interfaces, Human-Computer Interaction Series,
Springer London, 2010, Ch.~3, pp. 35--54.
\bibitem{VID10}
C.~Vidaurre, B.~Blankertz, {Towards a cure for {BCI} illiteracy.}, Brain
topography 23~(2) (2010) 194--198.
\bibitem{HAM12}
E.~M. Hammer, S.~Halder, B.~Blankertz, C.~Sannelli, T.~Dickhaus, S.~Kleih,
K.-R. M{\"u}ller, A.~K{\"u}bler, {Psychological predictors of {SMR}-{BCI}
performance}, Biological Psychology 89~(1) (2012) 80--86.
\bibitem{Obermaier2001}
B.~Obermaier, C.~Guger, C.~Neuper, G.~Pfurtscheller, {Hidden Markov models for
online classification of single trial EEG data}, Pattern Recognition Letters
22~(12) (2001) 1299--1309.
\bibitem{SHE06}
P.~Shenoy, M.~Krauledat, B.~Blankertz, R.~P.~N. Rao, K.-R. M\"{u}ller, Towards
adaptive classification for {BCI}, Journal of Neural Engineering 3~(1) (2006)
R13+.
\bibitem{Lenhardt2008}
A.~Lenhardt, M.~Kaper, H.~Ritter, {An Adaptive P300-Based Online Brain Computer
Interface}, Neural Systems and Rehabilitation Engineering, IEEE Transactions
on 16~(2) (2008) 121--130.
\bibitem{Bin2009AnOnlineMultiCh}
G.~Bin, X.~Gao, Z.~Yan, B.~Hong, S.~Gao, {An online multi-channel SSVEP-based
brain-computer interface using a canonical correlation analysis method},
Journal of Neural Engineering 6~(4).
\bibitem{ZHU11}
D.~Zhu, G.~Garcia-Molina, V.~Mihajlovi\'{c}, R.~Aarts, Online {BCI}
implementation of {High-Frequency} phase modulated visual stimuli, in:
C.~Stephanidis (Ed.), Universal Access in Human-Computer Interaction. Users
Diversity, Vol. 6766 of Lecture Notes in Computer Science, Springer Berlin
Heidelberg, 2011, pp. 645--654.
\bibitem{Lee2003}
H.~Lee, S.~Choi, {PCA+HMM+SVM for EEG pattern classification}, Signal
Processing and Its Applications, 2003. Proceedings. Seventh International
Symposium on 1 (2003) 541--544.
\bibitem{kalunga2013ssvep}
E.~Kalunga, K.~Djouani, Y.~Hamam, S.~Chevallier, E.~Monacelli, {SSVEP
enhancement based on Canonical Correlation Analysis to improve BCI
performances}, in: AFRICON, 2013, IEEE, 2013, pp. 1--5.
\bibitem{Lu2010Regular}
H.~Lu, H.-L. Eng, C.~Guan, K.~Plataniotis, A.~Venetsanopoulos, {Regularized
Common Spatial Pattern With Aggregation for EEG Classification in
Small-Sample Setting}, Biomedical Engineering, IEEE Transactions on 57~(12)
(2010) 2936--2946.
\bibitem{BLA08b}
B.~Blankertz, R.~Tomioka, S.~Lemm, M.~Kawanabe, K.~R. Muller, {Optimizing
Spatial filters for Robust {EEG} {Single-Trial} Analysis}, Signal Processing
Magazine, IEEE 25~(1) (2008) 41--56.
\bibitem{LOT11}
F.~Lotte, C.~Guan, {Regularizing Common Spatial Patterns to Improve {BCI}
Designs: Unified Theory and New Algorithms}, Biomedical Engineering, IEEE
Transactions on 58~(2) (2011) 355--362.
\bibitem{ANG12}
K.~K.~K. Ang, Z.~Y.~Y. Chin, C.~Wang, C.~Guan, H.~Zhang, Filter bank common
spatial pattern algorithm on {BCI} competition {IV} datasets 2a and 2b.,
Frontiers in neuroscience 6.
\bibitem{YAN12a}
Y.~Yang, S.~Chevallier, J.~Wiart, I.~Bloch, {Automatic selection of the number
of spatial filters for motor-imagery BCI}, in: M.~Verleysen (Ed.), {European
Symposium on Artificial Neural Networks (ESANN)}, 2012, pp. 109--114.
\bibitem{Hardoon2004}
D.~R. Hardoon, S.~R. Szedmak, J.~R. Shawe-Taylor, {Canonical Correlation
Analysis: An Overview with Application to Learning Methods}, Neural Comput.
16~(12) (2004) 2639--2664.
\bibitem{Lin2007}
Z.~Lin, C.~Zhang, W.~Wu, X.~Gao, {Frequency Recognition Based on Canonical
Correlation Analysis for SSVEP-Based BCIs}, Biomedical Engineering, IEEE
Transactions on 53~(12) (2006) 2610--2614.
\bibitem{absil2009optimization}
P.-A. Absil, R.~Mahony, R.~Sepulchre, Optimization algorithms on matrix
manifolds, Princeton University Press, 2009.
\bibitem{barachant2010riemannian}
A.~Barachant, S.~Bonnet, M.~Congedo, C.~Jutten, {R}iemannian geometry applied
to {BCI} classification, in: Latent Variable Analysis and Signal Separation,
Springer, 2010, pp. 629--636.
\bibitem{barachant2012multiclass}
A.~Barachant, S.~Bonnet, M.~Congedo, C.~Jutten, {Multiclass brain--computer
interface classification by Riemannian geometry}, Biomedical Engineering,
IEEE Transactions on 59~(4) (2012) 920--928.
\bibitem{panicker2010adaptation}
R.~C. Panicker, S.~Puthusserypady, Y.~Sun, Adaptation in p300 brain--computer
interfaces: A two-classifier cotraining approach, Biomedical Engineering,
IEEE Transactions on 57~(12) (2010) 2927--2935.
\bibitem{schettini2014self}
F.~Schettini, F.~Aloise, P.~Aric{\`o}, S.~Salinari, D.~Mattia, F.~Cincotti,
Self-calibration algorithm in an asynchronous p300-based brain--computer
interface, Journal of neural engineering 11~(3) (2014) 035004.
\bibitem{verschore2012dynamic}
H.~Verschore, P.-J. Kindermans, D.~Verstraeten, B.~Schrauwen, {Dynamic stopping
improves the speed and accuracy of a P300 speller}, in: Artificial Neural
Networks and Machine Learning--ICANN 2012, Springer, 2012, pp. 661--668.
\bibitem{barachant2012bci}
A.~Barachant, S.~Bonnet, M.~Congedo, C.~Jutten, et~al., {BCI} signal
classification using a riemannian-based kernel, in: Proceeding of the 20th
European Symposium on Artificial Neural Networks, Computational Intelligence
and Machine Learning, 2012, pp. 97--102.
\bibitem{BAR13}
A.~Barachant, S.~Bonnet, M.~Congedo, C.~Jutten, Classification of covariance
matrices using a riemannian-based kernel for {BCI} applications,
Neurocomputing 112 (2013) 172--178.
\bibitem{yger2013review}
F.~Yger, {A review of kernels on covariance matrices for BCI applications}, in:
Machine Learning for Signal Processing (MLSP), 2013 IEEE International
Workshop on, IEEE, 2013, pp. 1--6.
\bibitem{jayasumana2013kernel}
S.~Jayasumana, R.~Hartley, M.~Salzmann, H.~Li, M.~Harandi, Kernel methods on
the riemannian manifold of symmetric positive definite matrices, in: Computer
Vision and Pattern Recognition (CVPR), 2013 IEEE Conference on, IEEE, 2013,
pp. 73--80.
\bibitem{xie2013nonlinear}
Y.~Xie, J.~Ho, B.~Vemuri, On a nonlinear generalization of sparse coding and
dictionary learning, in: Proceedings of the 30th International Conference on
Machine Learning, NIH Public Access, 2013, p. 1480.
\bibitem{goh2008unsupervised}
A.~Goh, R.~Vidal, {Unsupervised Riemannian clustering of probability density
functions}, in: {Machine Learning and Knowledge Discovery in Databases},
Springer, 2008, pp. 377--392.
\bibitem{goh2008clustering}
A.~Goh, R.~Vidal, {Clustering and dimensionality reduction on Riemannian
manifolds}, in: {Computer Vision and Pattern Recognition, 2008. CVPR 2008.
IEEE Conference on}, IEEE, 2008, pp. 1--7.
\bibitem{PEN06}
X.~Pennec, P.~Fillard, N.~Ayache, A riemannian framework for tensor computing,
International Journal of Computer Vision 66~(1) (2006) 41--66.
\bibitem{barachant2013riemannian}
A.~Barachant, A.~Andreev, M.~Congedo, et~al., The {R}iemannian potato: an
automatic and adaptive artifact detection method for online experiments using
{R}iemannian geometry, in: Proceedings of TOBI Workshop IV, 2013, pp. 19--20.
\bibitem{amari2010information}
S.-I. Amari, Information geometry in optimization, machine learning and
statistical inference, Frontiers of Electrical and Electronic Engineering in
China 5~(3) (2010) 241--260.
\bibitem{samek2013robust}
W.~Samek, D.~Blythe, K.-R. M{\"u}ller, M.~Kawanabe, Robust spatial filtering
with beta divergence, in: Advances in Neural Information Processing Systems,
2013, pp. 1007--1015.
\bibitem{samek2014information}
W.~Samek, K.-R. Muller, Information geometry meets bci spatial filtering using
divergences, in: Brain-Computer Interface (BCI), 2014 International Winter
Workshop on, IEEE, 2014, pp. 1--4.
\bibitem{barachant2011channel}
A.~Barachant, S.~Bonnet, Channel selection procedure using {R}iemannian
distance for {BCI} applications, in: Neural Engineering (NER), 2011 5th
International IEEE/EMBS Conference on, IEEE, 2011, pp. 348--351.
\bibitem{barachant2010common}
A.~Barachant, S.~Bonnet, M.~Congedo, C.~Jutten, Common spatial pattern
revisited by {R}iemannian geometry, in: Multimedia Signal Processing (MMSP),
2010 IEEE International Workshop on, IEEE, 2010, pp. 472--476.
\bibitem{congedo2013new}
M.~Congedo, A.~Barachant, A.~Andreev, A new generation of brain-computer
interface based on riemannian geometry, arXiv preprint arXiv:1310.8115.
\bibitem{li2009eeg}
Y.~Li, K.~M. Wong, H.~De~Bruin, Eeg signal classification based on a riemannian
distance measure, in: Science and Technology for Humanity (TIC-STH), 2009
IEEE Toronto International Conference, IEEE, 2009, pp. 268--273.
\bibitem{li2012electroencephalogram}
Y.~Li, K.~Wong, H.~De~Bruin, Electroencephalogram signals classification for
sleepstate decision-a riemannian geometry approach, Signal Processing, IET
6~(4) (2012) 288--299.
\bibitem{lee2010introduction}
J.~Lee, Introduction to topological manifolds, Vol. 940, Springer, 2010.
\bibitem{jost2011riemannian}
J.~Jost, Riemannian geometry and geometric analysis, Vol. 62011, Springer,
2011.
\bibitem{moakher2005differential}
M.~Moakher, A differential geometric approach to the geometric mean of
symmetric positive-definite matrices, SIAM Journal on Matrix Analysis and
Applications 26~(3) (2005) 735--747.
\bibitem{fletcher2004principal}
P.~T. Fletcher, C.~Lu, S.~M. Pizer, S.~Joshi, Principal geodesic analysis for
the study of nonlinear statistics of shape, Medical Imaging, IEEE
Transactions on 23~(8) (2004) 995--1005.
\bibitem{fukunaga1990introduction}
K.~Fukunaga, Introduction to statistical pattern recognition, Academic press,
1990.
\bibitem{ledoit2004well}
O.~Ledoit, M.~Wolf, A well-conditioned estimator for large-dimensional
covariance matrices, Journal of multivariate analysis 88~(2) (2004) 365--411.
\bibitem{blankertz2011single}
B.~Blankertz, S.~Lemm, M.~Treder, S.~Haufe, K.-R. M{\"u}ller, Single-trial
analysis and classification of erp components—a tutorial, NeuroImage 56~(2)
(2011) 814--825.
\bibitem{schafer2005shrinkage}
J.~Sch{\"a}fer, K.~Strimmer, A shrinkage approach to large-scale covariance
matrix estimation and implications for functional genomics, Statistical
applications in genetics and molecular biology 4~(1).
\bibitem{pascal123theoretical}
F.~Pascal, P.~Forster, J.~P. Ovarlez, P.~Arzabal, Theoretical analysis of an
improved covariance matrix estimator in non-gaussian noise, in: IEEE
International Conference on Acoustics, Speech, and Signal Processing
(ICASSP)., Vol.~4, 2005.
\bibitem{pencina2008evaluating}
M.~J. Pencina, R.~B. D'Agostino, R.~S. Vasan, Evaluating the added predictive
ability of a new marker: from area under the roc curve to reclassification
and beyond, Statistics in medicine 27~(2) (2008) 157--172.
\end{thebibliography}