Classification of EEG data using

machine learning techniques

Martin Heyden

Department of Automatic Control

MSc Thesis
TFRT-6019
ISSN 0280-5316

Department of Automatic Control

Lund University
Box 118
SE-221 00 LUND
Sweden

© 2016 by Martin Heyden. All rights reserved.

Printed in Sweden by Tryckeriet i E-huset
Lund 2016
Abstract

Automatic interpretation of reading from the brain could allow for many interest-
ing applications including movement of prosthetic limbs and more seamless man-
machine interaction.

This work studied classification of EEG signals used in a study of memory. The goal
was to evaluate the performance of the state of the art algorithms. A secondary goal
was to try to improve upon the result of a method that was used in a study similar
to the one used in this work.

For the experiment, the signals were transformed into the frequency domain and
their magnitudes were used as features. A subset of these features was then selected
and fed into a support vector machine classifier. The first part of this work tried to
improve the selection of features that was used to discriminate between different
memory categories. The second part investigated the uses of time series as features
instead of time points.

Two feature selection methods, genetic algorithm and correlation-based, were im-
plemented and tested. Both of them performed worse than the baseline ANOVA
method.

The time series classifier also performed worse than the standard classifier. How-
ever, experiments showed that there was information to gain by using the time se-
ries, motivating more advanced methods to be explored.

Both the results achieved by this thesis and in other work are above chance. How-
ever, high accuracies can only be achieved at the cost of long delays and few output
alternatives. This limits the information that can be extracted from the EEG sensor
and its usability.

3
Acknowledgements

Thanks to Inês Bramão and Mikael Johansson at the Department of Psychology

Lund University, for answering my questions and making this project a possibility.

Thanks to everyone involved in the project at LTH: Bo Bernhardsson, Anders

Robertsson and Michelle Chong for their guidance and helpful feedback.

Thanks to Johan Eker at Ericsson for many good discussions.

Thanks to the Department of Automatic Control for giving me an inspiring work

environment.

5
Contents

1. Introduction 9
1. Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2. Project Partners . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3. The Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4. Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
5. Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
6. Report Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2. Brain Computer Interface Overview 15
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2. BCI as a Classification System . . . . . . . . . . . . . . . . . . 16
3. Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4. Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5. Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
6. Different types of BCI systems . . . . . . . . . . . . . . . . . . 23
7. State of the art . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3. Classification 25
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2. Example of some classifiers . . . . . . . . . . . . . . . . . . . . 27
3. Algorithm independent theory . . . . . . . . . . . . . . . . . . . 33
4. Support Vector Machines . . . . . . . . . . . . . . . . . . . . . 36
5. Time Series Classification using SVM . . . . . . . . . . . . . . 42
6. BCI classification . . . . . . . . . . . . . . . . . . . . . . . . . 44
4. Feature Reduction 46
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2. ANOVA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3. Correlation-Based Feature Selection . . . . . . . . . . . . . . . 49
4. Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 52
5. Experimental Setup 57
1. The Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2. Classification Setup . . . . . . . . . . . . . . . . . . . . . . . . 59

7
Contents

3. Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6. Method 63
1. Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2. Static Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3. Dynamic Classifier . . . . . . . . . . . . . . . . . . . . . . . . . 66
7. Results 68
1. Static Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2. Dynamic Classifier . . . . . . . . . . . . . . . . . . . . . . . . . 71
8. Discussion and Conclusions 72
1. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3. Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Bibliography 74

8
1
Introduction

1. Motivation
The human body is very capable of interacting with its surrounding. It can share
its thoughts by speaking and move around via muscles. In some cases this is not
possible due to disabilities or injuries. In other cases it would be beneficial if the
brain could interact directly with its surrounding without first having to go through
the rest of the body. For example activating an emergency button at the moment the
brain wants to could be faster than the person physically having to press it.

This might sound like science fiction but the technology is not that far away. One
way for the neurons in the brain to communicate is through electrical activity. Those
electrical activities will change depending of what the brain is trying to accomplish.
If that activity can be measured and interpreted, the intentions of a person could be
read without relying on the rest of his body.

One such method for measuring brain activity is electroencephalography (EEG). It

measures the mean membrane potential of populations of neurons through a series
of electrodes, usually attached to the scalp, see Figure 1.1. These devices have been
around for a long time and have been used to assist in the diagnosis of epilepsy
and sleep disorders by looking at the electroencephalographic activity. This has
already been aided by automatic classification of the data where it has been found
that certain EEG activities occur before the onset of seizures [Berg et al., 2010].
It can also be used to monitor the mental state of pilots [The pilot brain], pilot
quadcopters [LaFleur et al., 2013], control of humanoid robots [Chae et al., 2012]
or controlling exoskeletons for disabled persons [Schaap, 2016].

A system that reads signals from the brain and then interprets the signals to make
decisions or execute commands is called a Brain Computer Interface (BCI) and this
is an emerging research field.

9
Chapter 1. Introduction

Figure 1.1 A cap used for EEG measurements. The measurements are
read from a series of electrodes placed evenly over the head. Taken from
http://www.flickr.com/photos/tim_uk/8135755109/.

2. Project Partners
This project has been a cooperation between three entities: The Department of Psy-
chology at Lund University, the Department of Automatic Control at Lund Uni-
versity and Ericsson Research. Different goals were envisioned by each entitiy, but
it was believed that they could be combined into one project. These goals are de-
scribed below.

The psychology department is conducting research within memory research and

specifically, in the mechanism of memory creation and retrieval.

When someone is trying to remember something, the memory is said to be encoded

by the brain. When that memory is later being remembered, it is said to be recol-
lected. An interesting question is if the processes that happens during the encoding
of a memory is later replayed during recollection. One way of testing this is to see if
patterns present during encoding is also present during recollection [Jafarpour et al.,
2014].

There is currently no unifying theory for how this process work [Widmaier et al.,
2014].

The Department of Automatic Control is interested in using the EEG as a sensor in

robotics applications. It could also be used as a sensor in vehicles as a safety system
to monitor the state and intentions of the driver.

Ericsson is interested in finding new sensors and techniques that can be utilized with
the help of the new 5G technology. The EEG sensor could then be wireless and the
calculations done in the cloud. It could also be an interesting part of the internet of
things.

10
 3. The Data

Figure 1.2 Example of images used in the memory study used in this project. The pictures
come from three different categories: objects, faces and landmarks. The participants are asked
to try to memorize word/picture pairs and are later asked to recall the picture when presented
with the word. [Image source: Inês Bramão, Department of Psychology, Lund University.]

3. The Data
The data for this project was supplied by the Department of Psychology at Lund
University and was designed to study if patterns during encoding emerged during
recollection.

The study consisted of the participants trying to remember word-picture pairs. The
picture was from three categories: faces, landmarks and, objects. An example of the
pictures used in the study is depicted in Figure 1.2.

EEG measurements were recorded during the entire experiment. However, only the
recordings from memorization were used in this work. More details about the data
is given in Chapter 5.

4. Challenges
It is challenging to automatically classify EEG signals due to the problems with
the signals described below. Advances in computing power, signal processing, ma-
chine learning, and EEG reading devices have however increased the capabilities of
automatic interpretations of the results.

The Signals
Electroencephalography (EEG) uses electrodes placed over the skull to measure
electric activity in the brain.

A conductive gel is usually applied to the electrodes to improve reading quality. For
an every day BCI system this is not feasible. And even then the EEG measurements
are on the µV scale and have low signal-to-noise ratio. The signals are further cor-
rupted by artifacts, electrical signals originating from, for example, eye and muscle

11
Chapter 1. Introduction

30

20

10

0
µV

−10

−20

−30

−40
−1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
seconds

Figure 1.3 Plot of EEG readings from one of the channels during the memory experiment.
Notice that it is on the magnitude of micro Volts and that at least two different frequency
components can be seen.

movement. These artifacts might be much stronger than the electrical activity that is
studied. The signals also often have multiple different frequency components even
in the absence of artifacts. Another problem is the limited spatial resolution.

The strength of EEG lies in its high temporal resolution and that it is cheaper and
more portable than many other brain imaging techniques.

An example of the readings from channel 24 can be seen in Figure 1.3. It can be
seen that there are at least two different frequencies present. Notice also the low
amplitude of the magnitude of ten micro Volts.

Classification and feature selection

This work will focus on the problem of classification. The goal of classification is
to use measurements to decide what the source of the measurements is. Choosing
what measurements to use is a large part of the problem. The chosen measurements
are called features. The different sources are called classes.

When trying to classify between a mouse and an elephant there are many possible
measurements. One could measure the color of the animals presented but that would

12
 4. Challenges

Figure 1.4 Time-frequency transformation of EEG readings. The transformation is done

independently for each channels and the mean of the results is displayed. The stimuli is
presented at time zero.

probably not be a very good feature since they both are gray. Their weight would
however be sufficient to discriminate between the two classes of animals.

For harder problems multiple features must be used to do efficient classification. In

this thesis, frequency transformation of the EEG signals was used to classify what
category the user is trying to remember.

Frequency transformation is a method for finding out the magnitude of the different
frequency parts of a signal, much like the Fourier Transform, but with the magni-
tudes varying over time. A picture of the mean of all time-frequency transforms
from an EEG recording can be seen in Figure 1.4

The time-frequency transformation gives many features and the amount needs to
be reduced somehow to help the training of the classifier. This reduction could ei-
ther be done manually, where the designer of the classifier chooses which features
are relevant for classification, or automatically where the feature selection is data
driven. This work will focus on testing new methods for automatic feature selection.

The EEG readings are time series. There are two different ways to classify such
measurements. A static classifier uses each measurement within the time series as

13
Chapter 1. Introduction

a feature, and does not use the time information. A dynamic classifier instead uses
each time series as a feature. Both type of classifiers were studied in this thesis.

5. Goals
This thesis will focus on two tasks.

The first task was to get an overview of the state of the art of EEG classification and
understand what methods are being used. This work is presented in chapters 2 and
3. While some methods was described in detail there was no time for a thorough
theoretical handling of all methods being found. The interested reader can read
further in the references given for methods that were not described with enough
detail.

The second task was to design an EEG classifier for the encoding phase of the
memory study. First a classifier that was previously used in similar research was
implemented as baseline. Then two new classifiers using different feature selection
were then implemented with the goal of improving on the baseline.

Limitations
In a field as big as brain computer interface there had to be limitations on what
was explored. In this project there was only one feature type, power spectral density
via continuous wavelet transform. There was also only one classifier type, support
vector machines. These were used in previous work. Instead, only feature selection
methods were explored. The usage of an SVM is also motivated by its convexity.
Always finding a global maximum for the design of the classifier is very convenient
when testing feature selection methods.

There were many possible techniques to explore in this project. Some notable that
had to be discarded due to time constraint included outlier rejection or to weight
samples differently. Deep learning techniques were also not explored.

6. Report Outline
Chapter 2 gives an overview of the BCI field. Chapter 3 covers classification and
Chapter 4 covers the feature reduction method used in this work. Chapter 5 de-
scribes the experiment in detail and the framework used for classification. Chapter 6
describes the method used in this thesis and their results are presented in Chapter 7.
The final discussion and suggestion for future work is given in Chapter 8.

14
2
Brain Computer Interface
Overview

1. Introduction
A Brain Computer Interface (BCI) is a system that takes inputs from the brain, for
example EEG, and translates it to commands on a computer. This is usually done
by classification, where several commands are trained. The training consists of the
user executing the different commands. The goal is then to build a system which
uses this information to classify future commands.

This work will not look into model-based EEG classification. This has been done
for seizure detection [Chong, 2013] where the signals are much stronger than for
ordinary brain activity. It is not assumed to be impossible to use a model-based
approach for normal signals, but such possibilities were not explored.

Instead of making a new literature study, results will be presented from Hwang et
al. [Hwang et al., 2013] that reviews the literature from 2007 to 2011. It does not
include the most recent research but should still be sufficient for giving an overview
of the field.

An illustration of a complete BCI system can be seen in Figure 2.1. The first step is
signal acquisition where data is acquired, for example via EEG. Next features are
extracted from the data and commands are decided based on those features.

Memory classification is not a BCI system since there are no commands executed.
Everything up the the command execution is however the same, and the execution
of commands is generally not the hard part of building a BCI system. Memory
classification will hence be treated as a BCI system throughout this report.

15
Chapter 2. Brain Computer Interface Overview

Figure 2.1 An overview of an BCI system taken from [Thorpe et al., 2005]. The signal ac-
quisition might for example be from EEG signals. Features are then extracted and translated
to commands via, for example, machine learning.

2. BCI as a Classification System

The following properties and problems need to be considered for all BCI systems
that rely on classification [Lotte et al., 2007].

Noise and Outliers: BCI features are very noisy and might contain outliers due to
the poor signal-to-noise ratio of EEG signals (or other measurement meth-
ods).

High Dimensionality: Due to the many channels and potential features per chan-
nel the feature vectors for BCI systems are usually of very high dimension.
This can be handled by feature reduction or by a classification method that
can handle many features.

Time Information: BCI contains information over time. This could be handled
by concatenating feature vectors for different time points or classifying time
series.

16
 3. Sources

Non stationary: EEG signals often vary over time, and it might even be the change
that is interesting.
Small training sets: It is usually very time consuming to get large training sets.
This is usually not feasible since the test-person could get fatigued. Hence
the classifier method needs to work well even with a low number of training
data.

3. Sources
There are many methods for reading signals that measure the brain activity. EEG
is most commonly used and also the choice for this work. Hence it will be given a
more thorough description than the other methods. The sources can be split into non
invasive methods where readings are done from outside the skull and invasive where
the readings are done from inside the skull. Invasive methods give better readings
but should of course be avoided whenever possible due to the danger of them. Hence
almost all future BCI systems should be using non invasive methods.

[Hwang et al., 2013] found that in the years 2007-2011 32% of sources were inva-
sive, and 60% used EEG. The remaining 8 % included fMRI and MEG introduced
later.

EEG
Electroencephalography (EEG) uses electrodes placed on the skull to measure elec-
tric activity in the brain. The electrodes are placed according to the international
10-20 system, see Figure 2.2, so that different experiments more easily can be com-
pared. The analog signals are then amplified and sampled by an A/D converter for
further handling.

A conductive gel is usually applied to the electrodes to improve reading quality.

This can take over an hour, even for skilled users. For an every day BCI system
this is not feasible. Dry electrodes are however becoming available. A camera is
often used to record eye movement and helps in artifact removal. Such a camera is
typically not present on an everyday BCI systems, but the artifact removal could be
done without the help of the camera.

The strength of EEG lies in its high temporal resolution and its relative low cost
compared to other methods.
What do we measure? EEG measures electrical activity in the brain. One of the
ways for the neurons in the brain to communicate is through electrical discharge.
The current released from just one or a couple of neurons is too low to be mea-
surable using non-invasive measurements. A large group of neurons must fire in
synchronization for its activity to be detected by an electrode. Hence, the strength

17
Chapter 2. Brain Computer Interface Overview

32 – Right Mastoid
VEOG – Vertical Eye
Channel 1 2
HEOG ­ Horizontal
Eye Channel Gd
3 7
4 5 6
8 12
9 10 11

13 14 15 16 17

18 19 20 21

23 24 25
22
26

27 28 29
30
31

Figure 2.2 Electrode placement for the 10-20 systems as seen from above the skull. The
triangle in the upper part of the figure represents the location of the nose. The Electrode
are spread evenly over the scalp to get measurements from the entire brain. There are also
sensors measuring eye movement. [Image source: Inês Bramão, Department of Psychology,
Lund University.]

of the signal depends on the number of active neurons and their synchronization
strength [Widmaier et al., 2014].

The EEG signals have for long been categorized into different frequency bands
corresponding to different activities. For example, the alpha rythm of 8-13 Hz is
present when the subject is relaxed while the beta rythm (>14 Hz) is dominant
when the subject is alert. Higher frequency is generally an indication of increased
alertness [Widmaier et al., 2014]. Different parts of the brains are also known to be
active during different tasks.

Limitations The main limitation of non invasive EEG’s is its poor spatial resolu-
tion and its poor ability to measure activity deep inside the brain [Srinivasan, 1999].
It is also biased to certain neurons types [Murakami and Okada, 2006] and the sig-
nals are also distorted by the skull.

18
 4. Preprocessing

Other methods
Below follows a short introduction to other methods to measure brain activity.

MEG Magnetoencephalography (MEG) is a technique similar to EEG, but it reads

the magnetic fields induced by the current in the brain.
MEG typically has hundreds of channels. Furthermore its magnetic readings
are less distorted than the electric potential of EEG by the skull.
MEG requires shielding from magnetic fields, including the earth’s. This
means that it can only be used in controlled environments such as in clini-
cal applications.

fMRI Functional magnetic resonance imaging measures brain activity by detecting

changes in blood flow. It uses the principle that active areas in the brain have
increased blood flow. The blood flow is detected using a strong magnetic
field to align the oxygen nuclei and another field to to locate the nuclei. The
method has very high spatial resolution, pinpointing the location in 3d-space,
but has low time resolution.

Invasive methods Invasive methods involve taking measurements under the skull.
The benefit is that the signals do not get distorted by the skull. It is however
a very dangerous procedure and should only be used when necessary.

4. Preprocessing
The EEG signal comes with a lot of noise. Contracting muscles or even moving
the eyes will result in big changes in the signal. Because of this the signals need to
be cleaned up from such artifacts. In clinical applications this is usually done by a
combination of independent component analysis, which tries to separate the signal
into independent components [Lee, 1998], and visual inspection. For a BCI system
there is no time for manual intervention and the clean up needs to be automated.

Automatic artifact removal is necessary for a BCI system but will not be covered in
this work.

5. Features
There are many different possible features to use for classifying EEG signals. A
brief overview of methods used in the field will be given, but for details please see
the references given below. As this project only considered one type of features
there will be no discussion of the different features strength and weaknesses due to
time constraints.

19
Chapter 2. Brain Computer Interface Overview

Figure 2.3 Overview of features used in BCI research. [Hwang et al., 2013]

Frequency
The Fourier transform of a signal x(t) defined for all t ∈ R is given by

Z ∞
X( f ) = x(t)e−i2π f t dt. (2.1)
−∞

The spectrogram of x(t) is given by S( f ) = |X( f )|2 and serves as an estimation

of the magnitude of different frequencies. This method has two obvious problems.
Firstly, it requires data on an infinite time domain, and secondly, the frequency is
the same for all time points. These two problems are attempted to be solved by the
so called short time Fourier transform which uses a window w to locally calculate
an approximation of the Fourier transform.

Z ∞
X(τ, f ) = x(t)w(t − τ)e−i2π f t dt (2.2)
−∞

The window function is usually symmetric and of unit L2 norm. The spectrogram
now has a time component and is calculated as S(t, f ) = |X(t, f )|2 . For further treat-
ment see [Sandsten, 2016].

For a window function with support in a finite interval, infinite future samples are
no longer needed, but for calculations for time t there still must be measurements
up to time t + ∆ where ∆ depends on the support of the window function. Lower
frequencies require larger delays.

20
 5. Features

Since the data will be used in a computer there is no use for continuous frequency
measurements. The short time Fourier transform has a discrete counterpart

∞
X(m, f ) = ∑ x[n]w[n − m]e−i2πn (2.3)
n=−∞

There are many ways to calculate the frequencies of the signals. Some of those
methods will be briefly introduced here for completeness, but this work does not
attempt to find the best method or finding the best configuration for the method of
choice.
Multitapers Multitapers use several independent windows to estimate from the
same data segment. The final estimation is then the average of all windows which
reduces the variance of the estimation, see [Percival and Walden, 1993] for further
treatment.
Continuous Wavelets One limitation of the short time Fourier transform is that it
uses the same window function for all frequencies. Using a big window function
will lead to bad time resolution and a small window will lead to bad frequency
resolution. The continuous wavelet transform of x(t) is calculated using a wavelet
function ψs,t0 (t)

Z ∞
∗
W (s,t0 ) = x(t)ψs,t 0
(t) dt, (2.4)
−∞

where x∗ denotes the complex conjugate of x, t0 is the center of the windows and s
is the timescale. For some windows s = 1f but this is generally not true. The wavelet
is calculated from the mother wavelet

 
1 t − t0
ψs,t0 (t) = √ ψ0 . (2.5)
s s

Notice how the wavelet gets smaller for larger frequencies. See [Hramov et al.,
2015] for further discussion.

Autoregressive coefficents
An autoregressive (AR) model is used to describe a random process where the next
value depends linearly on some of the previous values, and a white noise e. For a
scalar AR model of order p this can be described as

p
x[t] = ∑ αi x[t − i] + e[t]. (2.6)
i=1

21
Chapter 2. Brain Computer Interface Overview

For further discussion and how to calculate the coefficents αi , see [Jakobsson,
2013]. For EEG this needs to be extended to be time varying and multivariate. For
an adaptive autoregressive process, αi (t) is a function of time. This was used in
[Schlögl et al., 1997] where it was argued that the random nature of the process
well described the random nature of the EEG data.

The AR coefficents can then either directly be fed into a classifier or used for fre-
quency estimation [Krusienski et al., 2006]

σe2
Ŝ(ei2π f ) = p . (2.7)
|1 − ∑i=1 αi e− j2π f |2

where σe is the variance of the white noise e.

Hjorth parameters
Hjorth parameters were introduced by Bo Hjorth [Hjorth, 1970]. They consist of
three quantities which can be defined in the frequency domain but which also can
be calculated in the time domain.

Activity: The variance, which is the same as the power, of the signal var(y(t)).
Mobility: Describes the mean frequency and is calculated as
v  
u var dy y(t)
u
t dt
mobility(y(t)) = (2.8)
var(y(t))

Complexity: Describes the change in frequency and is calculated as

 
dy
mobility dt y(t)
complexity(y(t)) = (2.9)
mobility(y(t))

Phase synchronization
Most frequency methods only look at the amplitude, ignoring the phase. AR coef-
ficients also ignore the phase of the signals. The phase information could be very
relevant since two parts of the brain having the same phase indicate that they are
cooperating [Brunner et al., 2006]. There are multiple measures of synchronization
but they will not be presented here.

Phase has not been popular in BCI research. This might partly be due to BCI be-
ing a very m alsoulti-disciplinary research field and phase features require more
mathematical involvement.

22
 6. Different types of BCI systems

An example of when it is used is [Daly et al., 2012] where functional connectiv-

ity is used. Functional connectivity is defined as communication between different
regions and is identified via phase synchronization.

Event-Related Potential
Event-related potential (ERP) is the change in potential that happens due to stimuli.
The most commonly used ERP is P300. P300 reacts to a visual stimuli and after
about 300 ms the ERP can be observed. A common setup for a P300 BCI system
is a matrix of example letters where the user focuses on the one he or she wants
to write. Flashing the row or column of the desired letter elicits a P300 while the
others do not. The P300 ERP can be detected by spectral methods [Fazel-Rezai et
al., 2012].

ERP is a common method in BCI research. However, it will not be discussed further
since it is far from the rest of the work.

6. Different types of BCI systems

BCI systems can be split into different categories. What follows is a summary from
[Nicolas-Alonso and Gomez-Gil, 2012]. The first distinction is between systems
that rely on external stimuli and those that do not.

Exogenous: Exogenous BCI system uses neural activity that originates from exter-
nal stimuli. This method have the advantage of requiring only one channel,
very little training and having a bit rate of up to 60 bit/min. It does however
require permanent attention to the stimuli, limiting its use in industrial appli-
cations. A P300-based BCI system is an example of an exogenous system.
Endogenous: Endogenous BCI does not rely on any stimuli. It instead relies on
the user self regulating the brain rhythm. This technique has the advantage of
being operated when the user wants to, and without having to focus on the
stimuli. It does however require a lot of training (weeks or months) where
the trainie has to learn to produce specific patterns. This is done via neuro
feedback, where the user is presented how close to a pattern he is. Multi-
channel EEG recordings are also required for good performance. The bit rate
is typically 20-30 bits/min. A system based on imagined movement is one
example of an endogenous system.

BCI systems can further be split into synchronous, which only analysis the signals
during pre defined time bins, and asynchronous, which always looks at the signals
seeing if there is a command present. Synchronous is much easier to implement, but
asynchronous offers much more seamless interaction.

23
Chapter 2. Brain Computer Interface Overview

Endogenous asynchronous is the most interesting system, but also the hardest!

7. State of the art

This work did not have time for a complete overview of the start of the art results.
Instead the current capabilities of the BCI technology will be highlighted by pre-
senting some results in finger tapping classification.

Three properties of a BCI system are of interest: its speed, its accuracy, and the
number of possible outputs. The bit rate, or information content the classifier is
able to achieve depends on these three properties. A high bit rate is wanted, but if
the accuracy of the system is too low, it will be unsatisfactory to use. No bit rate
calculation was done in this work.

Blankertz et al. [2006] achieved a mean accuracy of 89.5 % on imagined movement

for detection of continuous tapping over trials of 2 seconds. 128 EEG channels were
used which is a lot of channels. It did however require very little training with the
help of presenting the user with feedback. The highest bit rate was at 35 bits/min
with an accuracy of 98 % while one subject had 15 bit per min and another could
not achieve any BCI control, illustrating the variability in current BCI performance.

Ang et al. [2008] achieved an accuracy of 76.7 % on imagined taps on 2 second

intervals. This study used 118 electrodes.

The cited work above managed to achieve good accuracy, but at the cost of only
being able to detect continuous tapping, and only in segments of 2 seconds. This
would lead to very delayed control. Daley et al. [2011] tried to identify single imag-
ined taps for different time intervals. For intervals of 1 second an accuracy slightly
above 80 % was achieved. 0.5 second gave an accuracy around 63% and was the
smallest window that was significant above chance on a α = 0.01 level. This work
only used 19 channels.

It can be noted that high accuracy can only be achieved by low time resolution,
limiting the information content that can be extracted from the sensor.

24
3
Classification

1. Introduction
The goal of the classification in this work is to be able to distinguish between the
three different memory categories by using the EEG data. To do so, the computer
must be able to recognize the different patterns that arise in the measurements of the
electrical potential of the brain. This is a standard problem in the field of machine
learning.

The problem could be viewed as a decision problem. The machine is presented

with some data and is then expected to make a correct decision based on that data.
Below follows some examples of general decision problems to illustrate the concept
of decision problems.

• Given a black jack hand, should the player hit or stand?

• Given measurements about wind and temperature, what will the temperature
be in two days?

• Given a picture of a face, who is the person?

For a black jack hand, the decision could be made by understanding the mathematics
of blackjack. For the weather prediction, the prediction could be made explicitly by
models of the weather which have been learned from previous experience. For face
recognition, it is common that the system has a set of training pictures which it uses
to make a prediction when presented with a new picture. For all these examples it is
clear that the decisions require some prior information about the problem.

In this work, the data was EEG signals and the goal was to make some decision
based upon them. In general, that decisions could be where to move the cursor on
a screen or how to move a prosthetic limb. Specifically in the memory experiment,

25
Chapter 3. Classification

the decision that had to be made was which of the three categories the user has tried
to remember.

Old samples with known memory category will be used to classify new ones. This is
called supervised classification and is the only machine learning technique that will
be used in this work. Supervised means that the labels of the training data are known.
Classification means that the output of the learning system is discrete, for example
"hit" or "stand". For regression based learning the output is instead continuous, for
temperature prediction the temperature could take on any value. For unsupervised
learning the machine learning is presented with data samples with unknown label.
Part of the learning task is to cluster which training data that belong together, for
example which pictures that depict the same person.

Outline of the Classification Problem

The previous discussion is now formalized for the case of supervised classification.
The task is to use a set of old training samples X = {~x i } to classify new samples
~xnew . Each sample contains a set of features ~x i = x1i , ..., xni . Choosing a good feature
set is an important part of getting good classifier performance and will be discussed
further in the next chapter.

Classification is made by designing a decision rule that assigns each feature vector
~x to a class yi based on the samples in the training set. The decision rule can then
be used to classify both old and new samples. Classifying the training set gives an
accuracy called the training accuracy.

The goal of the decision rule is to find the structure of the classes. A good training
accuracy is important as it can not be expected to do better on new data than on
older data. But it is the accuracy on a previously unseen set, called the testing test,
that is the true measure of the classifier’s performance. That accuracy is often called
testing accuracy. Using a model that is too complex will often lead to good training
accuracy, but bad testing accuracy as the model finds the structure in the training set
instead of the class structure.

Sometimes there are parameters in the classifier that have to be chosen by the user.
These are sometimes chosen in a data driven fashion by splitting the data into three
sets: Training, testing and evaluation. The classifier is trained on the training set for
different parameter configurations and evaluated on the testing sets. This could be
done by searching on a grid in the parameter space or by using some more advanced
search technique. The classifiers are evaluated on the testing set and the one that
performs best is then tested on the previously untouched evaluation set to yield its
generalization performance. Note that if the accuracy of the testing set was to be
used the classifier would be evaluated on data it has already seen!

26
 2. Example of some classifiers

5 First order polynomial

Forth order polynomial

Voltage(V ) 3

0 1 2 3 4 5 6
Current(A)

Figure 3.1 Example of using a too complex model. Black dots denote training samples
and blue denote testing samples. Green line is first order polynomial and red line is fourth
order polynomial. The fourth order polynomial has perfect training accuracy but that does
not mean that it is better at predicting future data.

Example: Polynomial interpolation To illustrate the concepts above a similar

problem will be presented that should be known to most of the readers. Imagine
that given n measurements (xi , yi ) one wants to predict yi given xi . The variable x
could for example be the current in a circuit and y the voltage. It is known that as
long as the resistance is constant there will be a linear relationship between the two
quantities. Due to measurement noise and other disturbances the relationship will
not be perfectly linear, but instead on the form y = ax + e where e might for example
have normal probability distribution. If one tries to fit a first degree polynomial to
the data it will most likely be impossible to get a perfect fit and there will be some
error on the set used to calculate the line. If instead a polynomial of degree n − 1
or higher is used, the curve will perfectly fit the samples. The generalization per-
formance will however most likely be worse. The result of using a first and fourth
order polynomial on four data points can be seen in Figure 3.1. 

2. Example of some classifiers

Next some classifiers will be presented. The purpose is twofold, to further illustrate
the classifier problem and also to present some of the popular classifiers used in
EEG classification. Since none of these classifiers will be used in the experiment of
this project the section can be skipped at the reader’s discretion.

27
Chapter 3. Classification

Figure 3.2 Example of Nearest neighbor classifier. Blue and red are the two different
classes. Black indicate new samples that should be classified and the circle around them
show the closest neighbor. Note how old samples are used to classify new.

Nearest Neighbor Classifiers

The nearest neighbor classifier has a very simple decision rule. For each new sample
it assigns the label of its closest neighbor in the training set. See Figure 3.2 for a
visualization. The classifier is generalized to k-Nearest neighbor classifier where
each new sample is assigned to the class that have the closest k’th neighbors. This
classifier is less sensitive to samples that are different from the rest of its class,
called outliers [Duda et al., 2000].

Linear Discriminant Analysis

Linear Discriminant Analysis (LDA) is a two-step classifier that starts with a feature
transformation that transforms the data into a lower dimensional space where the
classes should be as separated as possible. Here only the feature reduction will
be covered. The reason it is covered is because its frequent use in brain computer
interface research and that it illustrates well how a classifier tries to seperate the
data. Only the two class LDA will be covered. For the general case see for example
[Duda et al., 2000].

The goal of LDA is to find the projection onto a line that is best for discriminating
the data. This can mathematically be described as y = ~w ·~x where ~x is the original
feature representation and y is the new scalar representation of the sample. The
goal is to find the weight vector ~w with unity magnitude that best separates the two
classes, see Figure 3.3 for an illustration. There are many ways of measuring the
separation of the classes. LDA uses the sample mean defined as

1
~mi = ∑ ~x (3.1)
ni ~x∈D
i

28
 2. Example of some classifiers

Figure 3.3 An illustration of projection of points onto two different lines. The solid circles
are the samples and the color indicates their class. The hollow circles indicates their projec-
tions. It can be seen that for one of the projection the two classes are separated while it is not
on the other.

where ni is the number of samples in class i and Di is the set of samples in class i.
For the projected points the sample means are

1 1
m̂i = ∑ y = ni ∑ ~w ·~x = ~w · ~mi
ni y∈Y
(3.2)
i ~x∈D i

due to the linearity of the dot product. The difference in the projected means is then

|m̂1 − m̂2 | = |~w · (~m1 − ~m2 )|. (3.3)

This difference can be made arbitrarily large by increasing ~w but that will just be a
scaling. This is why ~w was forced to be of unit magnitude. To get good discrimina-
tion the difference in mean should be large compared to the variance of each class.
To that end, define the within-class scatter

ŝ2i = ∑ (y − m̂i )2 . (3.4)

y∈Yi

The total within-class scatter is then simply ŝ21 + ŝ22 . A reasonable goal is to find the
~w that maximizes

29
Chapter 3. Classification

|m̂1 − m̂2 |2
J(~w) = . (3.5)
ŝ21 + ŝ22

There are now two tasks left: Finding the optimum ~w and finding the decision point
on the projection.

It would be good if J(·) depended explicitly on ~w. To find such a description define
the scatter matrices

Si = ∑ (~x − ~mi )(~x − ~mi )T , Sw = S1 + S2. (3.6)

x∈Di

ŝ2i can then be rewritten as

ŝ2i = ∑ (y − m̂i )2 = ∑ (~wt~x − ~wt ~mi )2

y∈Yi ~x∈Di
(3.7)
= ∑ ~w (~x − ~mi )(~x − ~mi )t ~w = ~wt S i~w.
t
~x∈Di

The sum of the scatters can then be expressed as

ŝ21 + ŝ22 = ~wt S w ~w. (3.8)

The difference in means can be written as

(m̂1 − m̂2 )2 =(~wt ~m1 − ~wt ~m2 )2
=~wt (~m1 − ~m2 )(~m1 − ~m2 )t ~w (3.9)
=~wt S b ~w,

where S b = (~m1 − ~m2 )(~m1 − ~m2 )t . Thus the expression for J(·) becomes

~wt S b ~w
J(~w) = . (3.10)
~wt S w ~w

To maximize J(·) ~w must satisfy

S b ~w = λ S w ~w, (3.11)

which is a generalized eigenvalue problem [Duda et al., 2000].

30
 2. Example of some classifiers

For the n category case the data is typically instead projected onto the n − 1 orthog-
onal directions that best discriminate the data, see [Duda et al., 2000].

Hidden Markov Models

In image classification there is no temporal component that could be used, but for an
image stream or for EEG readings it could be useful to use that there is information
about which order the measurements come in. Hidden Markov Models have proven
to be useful in such problems [Duda et al., 2000]. Hidden Markov models will not
be used in the experiments of this work, but it constitutes an alternative method to
classifying EEG data which the author think could be interesting.

What follows is a brief introduction of Markov Models and Hidden Markov Models.
It might be insufficient for readers that have never dealt with Markov models before
and if so there are plenty of books on the subject, see for example [Brémaud, 2013].
Markov Models A (discrete) Markov Model is a random process that contains
N states denoted S = {S1 , ..., SN }. The process is realized by one state being vis-
ited at each time point q = {q(1), q(2), q(3), q(4), q(5), q(6)} where q(i) is the
state at time i. A possible realization for a three state Markov model could be
q = {S1 , S1 , S3 , S2 , S1 , S2 , S2 }.

The transitions between the states follow the Markov assumption saying that the
state at time t + 1 is only dependent on the state of time t. The probability of a
transition from one state to another is then fully described by the set of transitional
probabilities

ai j (t) = P(qt+1 = S j |qt = Si ),

which gives the probability that the state changes from state i to state j at time t. If
all the transition probabilities are time independent, the Markov model is said to be
homogeneous. This assumption is very common. An illustration of a Markov chain
can be seen in Figure 3.4(a).

The initial state of the Markov model might also be random. The probability of
starting in state i is denoted πi = P(q1 = Si ). For simulation of the model, these
probabilities need to be estimated or decided otherwise.
Hidden Markov Models For the hidden Markov model the states are not directly
observable but they give rise to some observations O = {O1 , O2 , ..., On }. Each state
has its own random distribution for the observations that could be either discrete or
continuous. An illustration can be seen in Figure 3.4(b).

For discrete observations the set V = {vi } describes all possible observation values.
The distribution for each state’s observations is then described by B = {b j (k)},

31
Chapter 3. Classification

a11
a11 a22
a21

S1
a12 a13 S1 S2

a21 a31 a12

a32 b1 (1) b2 (2)
S2 S3
b2 (1) b1 (2)
a23 v1 v2
a22 a33
(a) Markov Model (b) Hidden Markov Model

Figure 3.4 (a) Example of Markov model, (b) Example of hidden Markov model. The
states Si are not directly observable, but we can see the observations vi .

where b j (k) = P(vk at time t|qt = S j ).

To summarize, the following is needed to design a hidden Markov model with dis-
crete observation space.

1. The number of states N.

2. The possible observations V = {vi }.
3. The transition probabilities A = {ai j }.
4. The distribution of the observations B = {b j (k)}.
5. The initial distribution πi = P(q1 = Si ).

There are three problems that need to be solved to use HMM for classification.
Those will be presented, but the solutions see [Duda et al., 2000], [Juang and Ra-
biner, 1991].

The evaluation problem: Given a model λ = (A, B, π) and a set of observations O ,

compute P(OO|λ ) in an efficient way. This is the probability that the model generated
the observations.

The decoding problem: Given a sequence of observations O and a model λ deter-

mine the most probable set of states q .

The learning problem: Given a set of observation sequences {O

Oi }, determine the
O|λ ).
set of parameters λ = (A, B, π) to maximize P(O

32
 3. Algorithm independent theory

S1 S2 S3

Figure 3.5 Example of a left right Markov model. Once a node has been left, it can not be
revisited.

Example: Speech Recognition We now try to illustrate how we can build a clas-
sifier using the solution to these problem by presenting an example from [Juang and
Rabiner, 1991].

Consider the classification of individual words. Assume that each word can be rep-
resented as a time sequence of M unique spectral vectors. This is done by mapping
each time segment to the closest spectral vector. So for each word there is a se-
quence of observations O . By using the solution to the learning problem a hidden
Markov model can be built for each word.

By using the decoding problem it is possible to examine how good our model seems
to be and make adjustments to the number of states N.

Finally, the solution to the evaluation problem can be used to classify a new word
by calculating which model that was most likely to have generated the test word’s
spectral vector. 
Different HMM For speech recognition it makes sense to allow the process to
return to a node that has already been visited. But for some process it might be
beneficial to force the process to move forward by only allowing state transitions in
one direction. This can be implemented with a so called left right model, see Figure
3.5 . This has been used in BCI research [Obermaier et al., 2001].

3. Algorithm independent theory

Introduction
In this section properties that are general for all machine learning algorithms are
discussed. The two interesting theoretical questions "Is there an inherently stronger
classifier" and "Is there a best feature representation?" are answered. Next, general-
ization performance is studied using the concepts of overfitting and underfitting.

Superior Classifier
No Free Lunch theorem As stated in the introduction the goal is to get good gen-
eralization performance. An interesting question is if there is any classifier that can

33
Chapter 3. Classification

be considered generally superior, or if there even is one that is superior to chance.

The answer to this question is given by the No Free Lunch Theorem and it states
that in the absences of assumptions all classifiers are expeced to have the same
performance. For a formal statement of it see [Duda et al., 2000].

Ugly Duckling theorem So there is no reason to lean towards any classifier be-
fore the problem is taken into account. What about the choice of features? The
answer is given by the Ugly Duckling theorem, which states “There is no problem-
independent or "best" set of features or feature attributes” [Duda et al., 2000].

Note that the implication of this theorem is only that we can not choose a feature
set before taking the problem into account.

Classifier Complexity
In the introduction it was said that having a too complex classifier could lead to bad
testing accuracy even though the training accuracy was good. This is often called
overfitting and can be simplified as having two different sources [Hawkins, 2004]:

• Using a model that is more advanced than it need to be. For example, if the
data can be separated linearly it might do more harm than good to use a
quadratic decision surface.

• Using irrelevant or too many features.

The first part can be counteracted when designing the classifier and the second part
is counteracted during feature selection. It is, however, important that the feature
selection is adapted to the classifier that is later going to be used.

Due to the No Free Lunch theorem no classifier, including a simpler one, can be
expected to be better before taking the problem into account. Focusing too much on
avoiding overfitting could lead to underfitting, which means both bad training and
testing accuracy. This can be explained by

• Using a model that is too simple

• Using too few features.

Clearly, avoiding overfitting might lead to underfitting and vice versa. This will be
formalized with the help of the bias-variance tradeoff in the following paragraphs.

34
 3. Algorithm independent theory

Bias and variance for regression Even though the rest of the report does not deal
with regression it will be used here since it is a clear illustration of the bias and
variance for a classifier.

The goal of regression is to estimate the function F(xx) : Rn → R with the help of
training data D = {(xxi , y)}. This gives the regression function g(xx; D) : Rn → R
which is the estimation of F(xx). The effectiveness of the regressor could be mea-
sured by its mean squared error averaged over all possible training data sets D.

h i
ED [(g(xx; D) − F(xx))2 ] = (ED [g(xx; D) − F(xx)])2 + ED (g(xx; D) − ED [g(xx; D)])2
| {z } | {z }
bias2 variance
(3.12)

It is noticed that the expected error has two terms, the bias and the variance. The
bias is how far off it is expected to be on average. It should be intuitively clear that
higher bias leads to lower accuracy without looking at the formula above. Variance
measures how much the accuracy of the guess varies. Clearly a large variation in
accuracy means that the total accuracy suffers.

Now all that has to be done is to lower the bias and variance. The problem is that
once a model has been chosen it is hard to decrease them both at the same time.
Making the model more complex will lead to lower bias but higher variance. While
making it less complex will lead to lower variance, but higher bias. In the poly-
nomial fitting example earlier the first degree polynomial has some bias and low
variance while the fourth degree polynomial has zero bias, since over any training
set the error will be zero, but high variance.

With this in mind the design of a classifier could be seen as a two step process.
First a classifier model has to be chosen. It is important to chose one that fits the
problem well to reduce the bias and variance. One can also try to get as many
training samples as possible to further reduce bias and variance. The second step is
finding the optimal complexity for the model which gives the best combination of
bias and variance, i.e., the best compromise between over- and underfitting.

If one, for example, chooses a hidden Markov model with sufficiently many states,
all training samples can be classified correctly but test performance will most likely
be bad. If one instead chooses too few states the model will perform badly both in
training and in testing.

See Figure 3.6 for an illustration of the concept.

For classification the error is instead a product of the bias and the variance [Duda
et al., 2000], but the same principles still apply.

35
Chapter 3. Classification

Figure 3.6 Illustration of the bias variance trade-off. A too simple model leads to bad train-
ing and test accuracy. Using a too complex model will give good training accuracy, but bad
test accuracy as the model finds all the patterns in the training set instead of the general pat-
terns. The goal of the designer should be to find the complexity that maximizes test accuracy.

4. Support Vector Machines

A support vector machine (SVM) uses a linear decision boundary to classify new
samples. The training samples are for now assumed to be linearly separable and the
case for which they are not will be handled later. There are typically many lines
that separate the data, see Figure 3.7. The idea of SVM is to choose the decision
boundary that maximizes the margin of the training data in the sense that the min-
imum margin is maximized. The minimum margin is the distance to the decision
boundary of the samples closest to it. The reasoning being that this will hopefully
maximize the chance for new data to be correctly classified.

A new sample ~xi can be classified according to

if ~w ·~xi + b ≥ 0 then positive, otherwise negative. (3.13)

So the orientation of line ~w and the bias b must be found. For the training samples
it is required that

36
 4. Support Vector Machines

Figure 3.7 An illustration of a support vector machine classifier. The solid black line is the
classification boundary that maximizes the margin of the training samples. The circled points
on the dashed line are the support vectors. They have the smallest margin to the decision
boundary. The orange solid line also separates the data, but its margin is worse.

(
~w ·~x+ + b ≥ 1
(3.14)
~w ·~x− + b ≤ −1,

where~x+ are positive samples and~x− negative samples. Equality will correspond to
the points with the smallest margin. There will be at least one vector of each class
that gives equality, otherwise the margin could be increased further. Those vectors
are called support vectors. The goal is to find the ~w and b that maximize the margin.
Introduce yi such that yi is +1 for positive samples and −1 for negative samples.
One can then rewrite (3.14) as

yi (~xi · ~w + b) − 1 ≥ 0 (3.15)

with equality for all support vectors. Taking any two support vectors of different
classes the margin can be expressed as

~w
margin = (~x+ −~x− ) · (3.16)
||~w||

Using (3.15) gives ~x+ · ~w = 1 − b and ~x+ · ~w = −1 − b. The margin can now be
written as

2 1
arg max margin = arg max = arg min ||~w|| = arg min ||~w||2 . (3.17)
~w ~w ||~w|| ~w ~w 2

37
Chapter 3. Classification

So 12 ||~w||2 should be minimized under the constraints in (3.15). This gives the La-
grangian which is to be maximized

1
L = ||w||2 − ∑ αi [yi (~w ·~xi + b) − 1], (3.18)
2 i

where αi ≥ 0 are the Lagrange multipliers. This will fall out as a quadratic opti-
mization [Cristianini and Shawe-Taylor, 2000] for which there are many solution
methods. There is one more nice property of the SVM left to discover. Differentiat-
ing with respect to ~w and b gives

( (
δL
δ ~w = ~w − ∑i αi yi~xi = 0 ~w = ∑i αi yi~xi
δL ⇒ (3.19)
δb = − ∑i αi yi = 0 ∑i αi yi = 0

Since αi is non zero only for support vectors [Cortes and Vapnik, 1995] the decision
boundary is only dependent on the support vectors which make geometrical sense.
Substituting the expression for ~w in (3.19) into (3.18) gives

1
L = (∑ αi yi~xi ) · (∑ α j y j~x j ) − ∑ αi yi~xi (∑ α j y j~x j ) − b ∑ αi yi + ∑ αi (3.20)
2 i j i j i i

1
= ∑ αi − ∑ αi α j yi y j~xi ·~x j (3.21)
2∑i j

and substituting ~w from (3.19) into (3.13) gives for classification of ~xnew

~w ·~xnew + b = ∑ αi yi~xnew ·~xi . (3.22)

i

Note that we only have to calculate the scalar product between the training samples,
both for training of the classifier (3.21) and classification of new samples (3.22).
If we were to map our data to another domain using a function Φ(~x) : Rn → RN
we would need to calculate Φ(~xi ) · Φ(~x j ) . This could be beneficial since data not
separable in one domain could be in another.

Example [Cortes and Vapnik, 1995]: Imagine we want to have a decision boundary
corresponding to a second degree polynomial. We would then from our original
features x1 , ..., xn create the following features z1 , ..., zN

38
 4. Support Vector Machines

z1 = x1 , ..., zn = xn : n dimensions
zn+1 = x12 , ..., z2n = xn2 : n dimensions
n(n−1)
z2n+1 = x1 x2 , ..., zN = xn xn−1 : 2 dimensions

We can see that for a second order polynomial the dimensionality of the new space is
much larger than the original one. Calculating Φ(~xi )·Φ(~x j ) =~zi ·~z j could potentially
take very long time. This can be avoided by the kernel trick. 

The idea is to replace Φ(~xi ) · Φ(~x j ) with a kernel function K(~xi ,~x j ) that never has to
map the data to the higher dimensional space. There is some Hilbert space theory
for when this is possible. Details will not be covered, but one theorem used later
will be stated.

Theorem: Mercer Condition [Cortes and Vapnik, 1995]. A necessary and suffi-
cient condition for that K(uu, v ) defines a dot-product in a feature space is that

Z Z
K(uu, v )g(uu)g(vv) duudvv ≥ 0 (3.23)

for all g such that

Z
g2 (uu) duu < ∞. (3.24)

For a polynomial of degree d it is known that K(~xi ,~x j ) = (~xi ·~x j + 1)d can be used
[Cortes and Vapnik, 1995]. This is much more efficient than mapping the vectors
first and then calculating the scalar product.

Radial basis function (RBF) uses the following kernel

K(uu, v ) = exp −γ|uu − v |2 ,

(3.25)

which gives desicions on the form [Cortes and Vapnik, 1995]

!
n
2
f (xx) = sign ∑ αi exp{γ|x − xi | } (3.26)
i=1

If γ is chosen big enough 100% training accuracy can be achieved as only local
information is then used. Smaller γ will mean that more points are used and the

39
Chapter 3. Classification

Figure 3.8 An illustration of a decision boundary for a radial basis function kernel. A linear
decision boundary is found in the kernel space which then maps to a non linear boundary in
the original space.

decision boundary will be less complex. An illustration of a decision boundary from

an RBF kernel can be seen in Figure 3.8.

The benefits of SVM is that it guarantees finding a global minimum and that it can
efficiently map the data to a higher dimension without paying with much higher
computation time. So far it was assumed that the data is linearly separable, at least
in some domain. Being restricted to finding a domain where the data is linearly
separable would make the method useless for most cases. Luckily this can be coun-
teracted.

Non separable case

For the non separable case we require that

(
yi (~xi · ~w + b) ≥ 1 − ξi
(3.27)
ξi ≥ 0

This allows for some of the samples to be on the other side of the margin or even on
the other side of the decision boundary. ξi is the distance from the margin. We then
minimize

40
 4. Support Vector Machines

1
min ||~w||2 +C ∑ ξik . (3.28)
2 i

We notice that choosing a large regularization parameter C will lead to fewer train-
ing errors but possibly lead to overfitting since training errors are heaviliy penalized.
The standard choices of k is 1 or 2, leading to the 1-norm and 2-norm soft margin
problem. For the solution to those see [Cristianini and Shawe-Taylor, 2000]. It is
clear that the 1-norm is less sensitive to outliers.

Even for the separable case it could be useful to allow for some samples to have
some error ξi to increase the margin. This could lead to lower training accuracy, but
also to higher validation accuracy.

Generalization Performance
The regularization parameter C gives the user good control with regards to over-
fitting. For radial basis function the user have further control with the parameter
γ.

SVM have good generalization properties and are insensitive to overtraining [Lotte
et al., 2007]. The complexity of a SVM is based on the decision boundary and not
the number of features [Joachims, 1998]. Hence for a linear kernel many features
can be used while still keeping a low complexity kernel. Using a non linear ker-
nel makes the model more complex and makes overfitting an issue that must be
considered.

SVM multiclass classification

The observant reader might have noticed that SVMs can only classify between two
classes. There are two standard methods to extended the SVM framework to multi-
class problem.

one against one In one vs one classification, each class is paired with all the other
classes. The final result is the class that wins the most duels. Ties can be broken as
long as the same features are used for all classifiers.

one against all In one vs rest each class is instead fed into a classifier problem
were the competing class consists of all other classes. Hopefully only one class
wins against the rest, but if multiple do, the resulting tie can be broken.

See [Hsu and Lin, 2002] for further treatment.

41
Chapter 3. Classification

0.5

−0.5

−1
0 1 2 3 4 5 6

Figure 3.9 An example of how the Euclidean norm might be insufficient to compare time
series. The black seesaw signal is closer to the green sinusoid than the red delayed sinusoid
is. If one wants to decide if the given signal is a sinusoid or a seesaw signal the time shift is
irrelevant.

5. Time Series Classification using SVM

It is possible to classify time series with SVMs using the theory covered so far by
using each data point from the time series as a feature. This could, however, be un-
satisfying since the time dynamics are not taken into account. If all time points were
shuffled randomly (same for all samples) the classification would work out exactly
the same. Clearly some information is lost in the shuffling, but it is important to re-
member that adding explicit time series is not guaranteeing improved performance.

This work will extend SVM to allow for classification of time series via dynamic
time warping (DTW) which will be covered next.

Dynamic Time Warping

There are many ways of measuring the similarity of two time series x = {xi } and
y = {yi }. One might be tempted to use the L2 norm |xx −yy|. This simple approach has
some problems. Two identical signals with just one sample delay could be measured
as very different. An illustration of this can be seen in Figure 3.9.

The idea of dynamic time warping is to counteract time shifts and change in time
scale by finding the alignment of the signals that has the lowest distance. This is
done by defining a warping between the first x = (x1 , x2 , ..., xN ) and the second
signal y = (y1 , y2 , ..., yN ). The warping is defined by the warping sequence p =

42
 5. Time Series Classification using SVM

(a) Valid Path (b) Invalid Path

Figure 3.10 (a) shows an illustration of a valid warping. (b) shows an invalid warping
where the invalid paths are marked red. The first one takes a too long step and the second one
does not satisfy the monotonicity condition.

(p1 , p2 , ..., pL ) where pl = (nl , ml ) means that xnl and yml should be compared. The
warping path must satisfy the following [“Dynamic Time Warping” 2007]

1. The boundary conditions states that p1 = (1, 1) and pn = (N, N).

2. Monotonicity condition n1 ≤ n2 ≤ ... ≤ nn and m1 ≤ m2 ≤ ... ≤ mn .
3. Step size pl+1 − pl ∈ {(1, 0), (0, 1), (1, 1)}.

An illustration of a valid path and an invalid path can be seen in Figure 3.10. Note
that a warping path might contain more comparisons than the L2 norm. The cost for
a path is then calculated as

L
c p (xx, y ) = ∑ c(xnl , yml ). (3.29)
l=1

The distance between two sequences is then chosen as one of the potentially multi-
ple minimum paths. Now, only the question of finding the optimum path p is left.
Calculating all possible paths would be very slow. This problem is solved by dy-
namic programming, which allows the calculation to be done in O(N 2 ) time com-
plexity.
Multivariate DTW To reach good classifier performance it would be beneficial to
use multiple time series. The two simplest ways to handle this would be to either
warp them independently or force them to all have the same warp.

43
Chapter 3. Classification

Normalization Normalizing each time point independently of the other time

points would defeat some of the purpose of the time warping. One possible nor-
malization is to normalize the time series, but then scaling information would be
lost.

Using DTW In Support Vector Machines

With the theory covered so far it would be possible to design a k-nearest neighbor
classifier using dynamic time warping as the distance measure. There is still some
work to be done to use it in a support vector machine.

One might be tempted to define a kernel as k(xx, y ) = −DDTW (xx, y ) but this is not
a valid Kernel[Gudmundsson  et al., 2008]. Neither is the gaussian kernel attempt
−DDTW (xx,yy)
x y
k(x , ) = exp σ2
[Lei and Sun, 2007]. The problem is that they are not
positive semidefinite which means they do not satisfy the Mercer condition, stated
in (3.23). This means that the optimization problem is no longer convex. While it is
still possible to find a separating hyperplane we can not be sure that it is the optimal
one. Cuturi [Cuturi, 2011] solved this by introducing the Fast Global alignment
kernel.

This work will use the pairwise proximity function SVM formulated by [Gud-
mundsson et al., 2008]. This method relies on a proximity function instead of a
kernel function. Let X be the feature space, then P : X × X 7→ R maps each pair of
feature vectors to their distance.

Given the training set {xxi } define

 
φ (x) = P(xx, x 1 ), P(xx, x 2 ), ..., P(xx, x n ) . (3.30)

This function maps the features to a new feature space where each feature is the
distance to one of the training vectors. P will be the DTW distance and the SVM
can be trained using these features.

6. BCI classification
There are two ways to handle the time property of BCI signals. When classifying a
time interval a static classifier just concatenates the features of different time points
into a feature vector for the whole time segment. Using this technique the informa-
tion at each time point is used. However, the information that time point t1 comes
before t2 is completely lost. This can be realized by concatenating the time vectors
in a different order. The classifier would work exactly the same!

44
 6. BCI classification

A dynamic classifier instead classifies a sequence of feature vectors, like hidden

Markov models classifiers. This is a common practice in speech recognition, but
not very used in BCI research.

In [Hwang et al., 2013] it was found that the dominant classifiers in BCI research
during the period 2007-2011 was Linear Discriminant Analysis (36%) and support
vector machines with 17%. The rest were below 5 %.

A classifer for a BCI problem often needs to handle noisy high dimensional data
and few training sets.

45
4
Feature Reduction

1. Introduction
Motivation
In some classification problems there is a lot of measurements that are candidate
features for classification. Reducing the amount of features fed into the classifier is
called feature reduction. A perfect classifier would have no need for feature reduc-
tion since it would be able to ignore irrelevant features. In practice no classifier is
perfect but there is still no guarantee that we can improve its accuracy with feature
selection. For EEG signals there are a lot of candidate features so it is likely that
a good feature selection can improve the accuracy by removing irrelevant features
and reducing overfitting.

The original feature input might also be so large that it has considerable impact on
the running speed, limiting real time performance. This further motivates the use of
feature reduction.

Overview
Feature reduction can be split into two categories, feature selection where only some
of the original features are kept and feature extraction where the original feature
space is transformed into one with lower dimension. Methods can further be split
into supervised, where the labels of the training samples are known, and unsuper-
vised where they are not known.

Feature selection is usually employed when there are many features to start with.
This means that the time complexity of the algorithm is very important. An exhaus-
tive search is generally not possible and one has to settle for a near optimal solution.
This is usually implemented by searching the feature space by a heuristic or non de-
terministic approach and finding a useful criterion to evaluate the current solution
candidate. Heuristic approaches tend to get stuck in local extrema, while non deter-

46
 2. ANOVA

ministic approaches need to be run multiple times as the results can differ between
runs due to its random nature.

The feature selection is done offline so while it is important to keep the speed rea-
sonably fast, it does not impact the run speed.

The two standard evaluation categories are wrapper methods and filter methods.
Wrapper methods is the intuitive choice since they use the classifier accuracy as
a metric. However, when an advanced classifier is used, the time it takes to train
the classifier for each feature set to be tested might be impractical. Filter methods
evaluate the feature set with some other measure. It can for example use correlation
or uncertainty measures from information theory [Huang, 2015].

One example of an heuristic seach apporach is Forward Selection. It starts with an

empty feature set and adds the feature that improves the evaluation function the
most. This contintues until a certain number of features have been added or there
are no features that improves the evaluation function [Rejer and Lorenz, 2013].
Backwards selection instead starts with all features and removes the feature that is
best to remove one at a time. Forward and backward selection can be combined so
that features are first added and then removed.

2. ANOVA
Analysis of variance (ANOVA) is often used when evaluating results of experiments
where different parameters are tested [Penny et al., 2011]. This work will use one-
way ANOVA, which tests the hypothesis that the means of different classes are the
same. This hypothesis is discarded on a confidence level α when there is a 1 − α
chance that the means are all not the same. So the chance that the hypothesis was
wrongfully discarded is α. It can also be used as a feature selection method by only
choosing features where the means are not all the same.

The data is assumed to be normally distributed for ANOVA. This is generally not
true for the EEG data. This does not mean that the tests are useless. A high confi-
dence level will still be a good indication that the means are not all the same.

ANOVA could be seen as a filter method, where the feature set is decided by each
feature being evaluated individually using the ANOVA test described above.

When there are only two classes present, a T-test is sufficient to decide if the means
are the same. One-way ANOVA is a generalization of this, so the theory for a T-test
will be presented first.

47
Chapter 4. Feature Reduction

T-test
For a T-test, the data is assumed to be drawn from two different distributions Xi ∈
N(µi , σi ). There are two competing hypotheses

(
H0 : µ1 = µ2
(4.1)
H1 : µ1 6= µ2

Using the measurements x1i and x2i the test quantity T can be calculated

x̄1 − x̄2
T=q (4.2)
s21 /N1 + s22 /N2

where x̄i is the sample mean, s2i sample variance and Ni the number of samples of
each class. The zero hypothesis H0 is discarded on significance level α if |T | >
t1−α/2 . Or put in other words, H1 is said to be true with probability 1 − α.

One-Way ANOVA
One-way ANOVA tests the hypothesis that all means are the same versus the hy-
pothesis that they are not all the same. So to be clear, it is enough that two means
are different.

Since only the three class case will be used in this work the presentation of ANOVA
will only use three classes. This is easily generalized to n classes.

Let the means be described by β1 , β2 , β3 and the number of samples for each class
by N1 , N2 , N3 . The samples are then assumed to be given by

yi j = βi + ei j (i = 1, ..., 3; j = 1, ..., Ni ) (4.3)

Where {ei j } is independent and N(0, σ 2 ) distributed. Then two quantities are cal-
culated

3
SSbetween = ∑ Ni (ȳi − ȳ)2 (4.4)
i=1

3 Ni
SSwithin = ∑ ∑ (yi j − ȳi )2 (4.5)
i=1 j=1

48
 3. Correlation-Based Feature Selection

where ȳ is the sample mean for all samples and ȳi is the sample mean for all the
samples in class i. SSbetween is a measure of the spread between the means of the
classes. SSwithin is a measure of the spread within the classes.

The test quantity for I classes is F = MSbetween /MSwithin , where

MSbetween = SSbetween /(I − 1) MSwithin = SSwithin /(n − I) (4.6)

where n is the total number of samples. The hypothesis that all the means are equal
are rejected at significance level α, by a comparison with the F distribution, if and
only if F > Fα;I−1,n−I . For a rigorous treatment see for example [Scheffé, 1959].

3. Correlation-Based Feature Selection

In the general sense two things are said to be correlated if knowing something about
one tells you something about the other. Clearly one wants features that are corre-
lated with the classes so that the value of the features can be used to make prediction
about the classes.

Correlation-based feature selection is a method for feature selection developed by

Hall [Hall, 1998]. It is based on the following principle

"Good feature subsets contain features highly correlated with the class, yet uncor-
related with each other."

The idea is that having features that are correlated adds very little information, while
increasing the risk of overfitting.

To penalize correlated features a feature subset S is evaluated by calculating

krc f
MS = p . (4.7)
k + k(k − 1)r f f

Where k is the number of features in the subset, rc f is the average class-feature

correlation and r f f is the average feature-feature correlation. A high feature-class
correlation and a low feature-feature correlation is wanted. So a larger value of MS
is better.

One might then ask what the point is of using correlation for feature selection.
Or rather in what scenarios it outperforms ANOVA. For features that separate the
classes into two groups, ANOVA will do a good job since their means will be differ-
ent. There could however be features that separate the samples into multiple groups

49
Chapter 4. Feature Reduction

Figure 4.1 An example where ANOVA would not recognize that the feature could be dis-
criminative. Red circles are one class and blue another. The classes have the same mean, but
their x value is still correlated with the class. Note that a linear decision boundary would not
do well in classifying the samples.

which are well separated, but the classes still have the same mean. In this case the
correlation measure could be better. For an example of a situation where the means
are the same but the classes still are well separated, see Figure 4.1.

Correlation Estimation
There are three methods proposed in [Hall, 1998] for correlation estimation: relief,
minimum description length and symmetrical uncertainty. Symmetrical uncertainty
was used in this work due to its ease of implementation. The others will not be
explored due to time constraints. All methods require the data to be discretized.
Method for this will be covered later.

Symmetrical uncertainty uses the entropy measure which is a measure of the uncer-
tainty of the system, hence the latter half of the name. The entropy of X is given
by

H(x) = − ∑ p(x) log p(x),

x∈X

where base 2 is used for the logarithm henceforth. The minus sign makes H a pos-
itive quantity since the logarithm of a probability is always negative. If p(x) = 0
then p(x) log(x) is defined as zero, which is consistent with the limit. Entropy can
be seen as an information measure. If p(x) = 1 for x = ξ then x only takes on that
value and no additional information is gained when one is presented with x, as it
was already known that x = ξ . Conversely if p(x) = 1/n for all x the entropy, or
information gained when presented with x is maximum. To see that the uncertainty
and the information measure are consistent with each other, note that the uncertainty
of a variable is the highest when the information gained when presented with it is
the highest.

Next, we define the entropy of X given Y . This is the information gained when
presented with X if Y is already known. If X and Y are independent this will be the
same as the entropy of X as no new information is supplied by Y . If, on the other
hand, the information about Y changes the probability of X the information could
be changed. This is mathematically defined as

50
 3. Correlation-Based Feature Selection

H(X|Y ) = − ∑ p(y) ∑ p(x|y) log p(x|y)

y∈Y x∈X

There are two cases. One where Y is another feature in which case the conditional
probability can be written as

p(x, y)
p(x|y) = .
p(y)

Clearly the distribution of a good feature will change when the class is known.

The other case is when Y is one of the classes. Then the conditional probabilities
will easily be calculated as the probabilities for the features of the samples within
that class.

If the entropy for X decreases when Y is supplied then additional information was
gained. Define this gain, also called mutual information, as

gain = H(X) − H(X|Y )

= H(Y ) − H(Y |X) (4.9)
= H(Y ) + H(X) − H(X,Y ),

where it is noticed that the last form can be used for calculation. Normalization can
be done so that the gains are comparable for different features via

 
gain
symmetrical uncertainty coefficient = 2 · .
H(X) + H(Y )

The factor two makes the range to be [0, 1].

Discretization
Two simple methods for discretization is equal interval width and equal frequency
intervals. Equal interval width splits the data into n bins where each bin covers an
interval of size l so that all features are covered. Equal frequency interval splits the
data into n bins where each bin contains the same amount of data points.

There are more advanced methods. One based on minimum description will be
briefly introduced below. For a full treatment see the original article [Fayyad, 1993].

51
Chapter 4. Feature Reduction

The idea works iteratively by finding a cut point on a segment. The cut should be
such that it minimizes the class entropy of the two new partitions. The entropy of a
segment S with k classes is given by

k
H(S) = − ∑ P(Ci , S) log(P(Ci , S)). (4.10)
i=1

where P(Ci , S) is the proportion of class i on S. The cut point T is chosen such that
it minimizes

|S1 | |S2 |
H(T ; S) = H(S1 ) + H(S2 ) (4.11)
|S| |S|

The computation speed is improved by noting that cut points can only be between
two classes. They further derive an expression for when keeping a cut based on the
minimum description length principle. Basically a cut is kept if the information paid
by doing the cut is repaid by better separation of the data.

4. Genetic Algorithm
The Genetic algorithm (GA) is, as the name suggests, inspired by the evolution in
nature that leads to fitter and fitter individuals. First the basic algorithm is introduced
[Whitley, 1994]. Then some possible variations are presented.

Motivation
The previous two methods only looked at each feature’s goodness independently.
The main motivation for using a genetic algorithm is its ability to evaluate the fea-
ture set as a unit. This is not useful for linearly separable features, but it could be ef-
ficient in finding features that only hold discriminative power when combined with
another feature. See Figure 4.2 for an example where the x or the y direction have
no discriminative power on their own but can classify the data perfectly together.

The Standard Genetic Algorithm

Introduction In nature the genome of populations is changing because fitter indi-
viduals have a higher chance of survival and subsequent reproduction compared to
those less fit. This makes the populations strive to be fitter. The genetic algorithm
tries to achieve the same thing. Now the fitness of the population is explicitly stated
as the fitness function f (X̄) and the goal is to find the best X̄ = (X1 , ..., Xn ). For an
optimization problem f (X̄) would be the function to optimize and X̄ the variable
for the function.

52
 4. Genetic Algorithm

0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1

Figure 4.2 An example where the two features have no disciminative power on thier own,
but do together. The blue circles and red crosses are the two different classes and the features
are the x and y coordinates. Note that no linear decision boundary can do better than chance.

In this work GA will be used for feature selection where the number of features
to be chosen are predetermined. The goal is to choose the n best features from the
feature candidates numbered 1, ..., N. The variable Xi can take on values in the range
[1, N] to represent the feature chosen, but duplicates are not allowed.

If the size of the feature subsets is allowed to change then the encoding need to be
binary and X̄ = (X1 , ..., XN ) and Xi = 1 encodes that feature i is present and Xi = 0
that it is not.

The fitness function will measure the goodness of the feature subset. When classifier
performance is used the method is a wrapper method. It is also possible to use some
other measure, making it a filter method.

The algorithm The algorithm is initialized with a genome of individuals. Each

individual consists of a vector that encodes that individual’s vector X̄, This vector
is often called a chromosome.

The algorithm then iterates by evolving the genome. The first step to this is calcu-
lating the evaluation function which is an absolute measure of how good the current
chromosome is, for example the value of f (X̄) or the accuracy of a classifier. Next

53
Chapter 4. Feature Reduction

the fitness function is calculated which is the chromosomes relative strength com-
pared to the rest of the genome. This can for example be calculated as fi / f¯ where
fi is the evaluation function of the chromosome and f¯ is the mean value of all the
evaluation function.

The chromosomes next go through the selection process, which yields an interme-
diate generation where a chromosome with a high fitness score might be present
multiple time and one with a lower might not be present at all.

Next is the recombination step, where there is a chance that each chromosome is
combined with another chromosome generating two new offsprings. This is called a
crossover and might, for example, be implemented by choosing a random crossover
point and splitting the chromosomes at that point and combining one half from each.

As a final step there is a small chance that an offspring goes through a mutation
where one of its binary values will flip. This is the end of the iteration, and is fol-
lowed by a new calculation of the fitness value. A graphical overview of the process
can be seen in Figure 4.3.

There are many possible stopping criteria. Examples include terminating after a
specific number of iterations, when the change in fitness is below a threshold or
when the maximum fitness is above a threshold.

Overfitting in GA When using the wrapper GA there will be a set of samples we

can use for training and evaluation. There is a risk that the resulting classifier gets
very well-tuned for the data that is presented to it. This is not overfitting in the strict
sense that was talked about in classification but the term is extended to cover this
case as well. The classical overfitting is still a potential problem if the size of the
feature subset is not prespecified. This can be counteracted by punishing genes with
many features [Duda et al., 2000].

The tuning problem can be counteracted by not presenting the same data to the
fitness function for each generation so that the classifier learns the trends and not
the testing set. Liu and Khoshgoftaar [Liu and Khoshgoftaar, 2004] showed that
random sampling technique (RST) reduces overfitting. The usage of RST means
that only a subset of the training set is used for each generation. Thus presenting
different patterns every time.

Why does it work?

There are many different variations and parameter sets for GA and there are no
strong theoretical statements for why it works. Instead, we have to believe in the
previous results shown, for example in the BCI field [Rejer and Lorenz, 2013], and
its good properties.

54
 4. Genetic Algorithm

selection

Initial fitness Crossover

Population function & mutation

Population

fittest
individual

Figure 4.3 A graphical illustration of the genetic algorithm. First an initial population is
chosen. Then the fitness value is calculated for each individual. This fitness value gives the
probability for each individual to survive the selection process. Those that survive the se-
lection process goes through crossover where 2 genes are combined together and mutation
where there is a probability that one feature might swap to another. The fittest individual is
extracted when the algorithm terminates.

Extensions
Diversity The diversity of a population in GA is a measure of how spread out
the genome is. Failing to maintain a diverse solution might mean that the feature
space is not adequately searched, while too high diversity could lead to very slow
convergence [D. Gupta, 2012].

One method of maintaining diversity of the population is to use a ranked space

where each individual is ranked both on fitness and diversity. The fitness function is
then calculated based on this rank. An overview of other methods can be found in
[D. Gupta, 2012].

Elitism In nature the parents can not live on for ever but in GA they can. Allowing
the parents to fight with their cross-over children for survival into the next gener-
ation means that a gene is never replaced with one that has worse fitness during
recombination. This might rapidly increase convergence, but at the cost of lower
diversity.

Parameter Control For the classic genetic algorithm the crossover rate and muta-
tion rate are set to be constant. The genetic algorithm is a very dynamic process so
it is reasonable to use dynamic parameters. It might, for example, be better to have

55
Chapter 4. Feature Reduction

higher mutations rate in the beginning to help explore the search space and lower
mutation rate later to fine tune the solutions.

Parameter control can be classified as deterministic in which it changes determinis-

tically for each generation, as adaptive where it changes depending on the properties
of the entire population or and self adaptive where the parameters themselves are
encoded in the genome and undergoes crossover and mutation. For a thorough rep-
resentation see [Eiben et al., 1999].

56
5
Experimental Setup

1. The Study
The data for this project was supplied by the Department of Psychology at Lund
University. The experiment is motivated by their research in memory representation.
A visualization of the setup can be seen in Figure 5.1.

The study was conducted in a Faraday cage at the department of Psychology at

Lund University. The subjects were Swedish native speakers without any known
neurological or psychiatric disorders. Both males and females were participating.

The stimuli material of the study was 64 different pictures from the three categories
"familiar faces", "landmarks" and "objects", yielding 192 pictures in total. The study
was split up into three parts. The first part was the familiarity task where the partic-
ipant is shown a picture on a computer screen for a couple of seconds. The pictures
were taken from one of the three categories. The participants were then asked how
well it represents the concept of the picture. This could in theory be used to remove
pictures that were not recognized by participants, but was not used in practice.

The second part was the study phase for which 24 word-picture pairs were pre-
sented and the participants was asked to try to remember the pair and also rate the
association between the pair.

Finally for the test phase the participant was shown a word and needed to retrieve
the category of the picture it was associated with. If they were correct they were
then also shown the original and a mirrored picture and had to choose which one
they thought they saw earlier. This was to force the participants to visualize the
picture, otherwise the retrieval is only conceptual. The goal of this was to increase
the classifier accuracy.

The study took approximately two and a half hour with breaks between each part.

The goal of the study at the psychology department was to study if any of the pat-

57
Chapter 5. Experimental Setup

Figure 5.1 Illustration of the experiment from which the data of this work originates. The
study consisted of three phases. During the familiarity phase the participants were showed
pictures from three categories:faces, objects, and landmarks. During the study phase they
were asked to memorize word/picture pairs. For the test part the participants were tasked with
remembering the picture that corresponded to the displayed word. Only the study phased was
used in this work. [Image source Inês Bramão Department of Psychology Lund University.]

terns during the encoding of the study phase were replayed during the recollection
of the test phase.

58
 2. Classification Setup

The data
EEG was measured throughout the entire experiments with a frequency of 2048 Hz.
This high frequency was used since it was necessary to detect triggers that were
sent each time the picture changed. The data was later down-sampled to 512 Hz.
The average of the left and the right mastoid is used as a reference for the data. That
means that all other measurements are relative to the reference.

For each part one EEG segment was saved. For the familiarity part data was saved
for the display of the picture. For the study part the data was also saved for the
display of the picture. Finally, for the test part data was saved for the display of the
word.

Data was saved from 1 second before the stimuli and up to 2.5 second after the
stimuli. The reason for saving data before the stimuli and for so long after was to
allow for frequency calculations. One such segment is called a trial.

Artifacts coming from sources such as eye and muscle movement were removed
using independent component analysis and the expertise at the psychology depart-
ment. The removal of eye artifacts was also helped by measurements from VEOG
and HEOG channels that measured vertical and horizontal eye movements. Trials
that could not be cleaned up to a satisfactory degree were removed.

2. Classification Setup
Features
Power spectral density will be used as features throughout the experiments. The
data to be classified will be split into time points. For each time point the magnitude
of a set of frequencies will be calculated. The phase of the signal will not be used.
This will be done for each channel yielding a three dimensional space to choose
features from, see Figure 5.2 for a visualization.

For a static classifier, features can be chosen arbitrarily over this three dimensional
grid. One might for example choose channel four, frequency magnitude at five Hz
and time point four as a feature. If one instead wants to use a dynamic classifier the
features must instead be time series. An example of a feature is then channel four,
frequency magnitude at five Hz and measurements from time point two to time point
eight.

Time bins
The goal of the study at the psychology department was to explore if part of the
process of remembering (encoding) a memory was replayed during the retrieval of
that memory. Those processes clearly will not be the same all the time. If one was

59
Chapter 5. Experimental Setup

Channel
number

Frequency

Time points

Figure 5.2 Illustration of the possible features to choose from. For each channel the mag-
nitude of some frequencies are calculated at different time points.

to build a classifier for the entire signal it would be impossible to tell at what time
eventual similarities happened.

To solve this, different classifiers are trained for different time intervals, called time
bins. A time bin starts at t0 and ends at time t1 and uses only data from that time
period for training. The training data is split into time bins that cover the whole trial
(could be overlapping). See Figure 5.3 for an illustration. All those time bins are
then used to classify the test phase. A classifier trained for a time bin that starts at t0
and ends at t1 for the study data needs not to classify the same interval for the test
data. The interval it is run on needs, however, to have the same interval size.

The time bin technique is useful also in other BCI applications. Waiting for the
whole signal before doing any classification would however lead to systems with
large delays. Daly et al [Daly et al., 2011] investigated the effect on accuracy for
different window sizes when identifying single taps. Smaller windows increase the

t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12

time bin 1 time bin 2 time bin 3

Figure 5.3 Illustration of the time bin concept. The samples are split into several indepen-
dent time bins containing some of the samples each.

60
 3. Previous Work

theoretical bit rate but at the cost of decreased accuracy.

The cited work also used sliding time windows. With this technique the output of
the classifier can switch with every time point instead of the length of the time
window, allowing the classification output to change more often without loss of ac-
curacy. Sliding time bins require to use some information from prior time bins to be
efficiently implemented, but in theory each time bin can be handled independently
of other time bins.

3. Previous Work
Jafarpour et al. [Jafarpour et al., 2014] studied at what point in recollection the
encoding of the memory was replayed. The study consisted of 3 phases. During
the first phase the participants were shown pictures of faces or scenes, which they
should not have seen before. If they had, the picture was removed from the exper-
iment. One classifier for faces and one for scenes were then trained from the data
that was kept.

For the second phase the participants were tasked with memorizing word-picture
pairs. For the last phase the participants were shown a word and asked to try to
remember the face/scene it was associated with. The classifier built was then used
during the third phase with the idea being that a high classification score during a
time slot meant that the original memory was replayed.

The machine learning used MEG readings from 274 channels. The data was then
transformed to the time frequency domain using 5 cycles Morlet wavelets (contin-
uous wavelet transform). The frequency scale went from 8-45 Hz with a resolution
of 1 Hz. Analysis was performed in time bins of 66 ms and there were 21 time
points in each time bin. Data was normalized by z-scoring (data normalized to zero
mean and unite variance) at each time point, frequency and channel across trials. A
support vector machine was the choice of classifier.

First they looked for category specific patterns during encoding. This was done by
using 10 fold cross validation. This means that for each time bin, 10 different clas-
sifiers were trained. Each classifier uses a different 10 % of the data for validation.
The accuracy was then the mean of the accuracy for the 10 classifiers. For each
validation pass a feature selection was run by only keeping those features that were
significantly different between the two categories in a T -test with 0.05 level. The
performance was then calculated as the average of all the 10 classifiers. The classi-
fiers that were significantly different from zero were then trained on all the data in
their time bin.

These classifiers were then used to check if the pattern present during their time
bin was later replayed during encoding which there were no indications of in the

61
Chapter 5. Experimental Setup

classifier performance.

They then checked if category specific patterns emerged during recollection. This
was done by running the classifier trained on data from the encoding time bin cen-
tered around 180 ms on the recollection trials where the participants correctly rec-
ognized the word cue. This gave significant classifier performance for the time bins
centered at 446 and 513 ms. Sliding time bins were not used. This could possibly
give better performance as there is no guarantee that the replayed pattern is aligned
with one of the time bins.

Finally, only those recollection trials where the participants also chose the correct
picture were tested. Now the classifier was only significant for 513 ms.

62
6
Method

1. Software
The matlab toolbox Fieldtrip toolbox [Oostenveld et al., 2011] was used in this
work. It is developed by Donders Institute for Brain, Cognition and Behaviour in
Nijmegen, the Netherlands, and have multiple methods for analysis of EEG and
MEG data. Methods for filtering the signals and to do time frequency calculations
using continuous wavelets were used.

For SVM classification LIBSVM [Chang and Lin, 2011] was used. It is written
in C++ and contains a Matlab interface. It supports one-vs-one multiclass SVM
classification.

2. Static Classifier
Introduction
The goal of this part was to explore alternatives to the ANOVA + linear SVM
method that was used by Jafarpour [Jafarpour et al., 2014] and by the psychology
department. It was decided to keep SVM as the classification method so the focus
could be kept on the feature selection. Other than ANOVA, two feature selection
methods were tested: correlation based feature selection and genetic algorithm.

Experiment
As the goal was to test which method performed best, it was important to have a
well defined goodness measure. Time bins is a good tool for memory research, but
it is hard to find a satisfying goodness measure for it. One could take the mean
accuracy over all time bins but it is not apparent that all time bins should be equally
important. It is yet unknown in what time segments good classifier performance is
expected from physiological reasons.

63
 Chapter 6. Method

Training Samples Training Samples

EEG {time, freq} features Output

Signal Processing Feature Selection Classification
Wavelet Transform •ANOVA Support Vector Machine
•Genetic
•CFS

Figure 6.1 Illustration of the classification pipeline used in this project. The EEG signals
are first transformed using wavelet transform. Then features are chosen from the possible
combination of time and frequency measurements. Those features are fed into a classifier
that outputs the prediction of the category for the input.

It was decided that it was better to evaluate classifier performance by classifying

the whole time segment at the same time. This could be seen as having one large
time bin. Different methods might be better at classifying different length of signals
(more efficient at using less data) or more efficient at different parts of the data.
However, this approach should serve well as a first evaluation method. If multiple
methods are close then further experiments should be conducted.

The classification pipeline of this work have three steps. The first is signal process-
ing, where the original EEG signals are transformed into the frequency domain. The
second step is feature selection, where the best features were chosen. Those features
are fed into the third step, which is classification. An illustration of the this can be
seen in Figure 6.1.
Base feature set For all experiments in this work the base feature set was the
frequency magnitude given by continuous wavelet transform.
Comparison with other work Using all the data at the same time instead of work-
ing with only a subset of it will mean that the accuracy will be at least as good as
the accuracy for the best time bin, and most likely better. This means that it is hard
to compare the results with other studies. But since there are no other published
results on this study that would be impossible anyway. Instead the comparison will
be between the old method, ANOVA, and the new tested in this work.

Rejected methods
Two simple methods were tested initially and rejected due to bad performance. Here
follows a short description of them
Principal Component Analysis Principal component analysis (PCA) finds the di-
rections were the data has the highest variance, see [Duda et al., 2000]. PCA was
used to find the directions where the spectral features had the largest variance. These
features were then fed into a linear support vector machine. The initial results were
not good enough to continue perusing this method.

64
 2. Static Classifier

This was not unexpected as there likely is large in-class variance in the data.
Linear Discriminant Analysis Linear discriminant analysis was used in a smiliar
way as PCA, i.e. projecting data into some new directions.. The method only gives
two directions for the three class problem. This was not sufficient to achieve good
classifier performance and was not further investigated.

ANOVA
Method The ANOVA method evaluates each feature independently and performs
a one way ANOVA test (Chapter 4.2). There are then two alternatives, either the n
most significant features are kept, or all features significant on an α level.

Next the features are normalized using z-scoring, so that all have zero mean and unit
variance. This is important so that the SVM does not give higher weight to features
with larger values.

For classification a linear SVM was used.

Motivation The usage of a linear kernel is motivated both theoretically and by
experiments.

The features chosen by ANOVA are different in means between at least two of the
three classes. So in that feature’s direction a linear separator should do a good job.
In the other directions there is no information. If a non linear kernel was to be used
the amount of features would have to be reduced significantly to avoid overfitting.
There is no theoretical motivation to do so since there is no indication that the
features would be good in a non linear kernel.

By experiments it was found that the previous arguments indeed were correct.

Correlations-Based Feature Selection

The correlation based feature selection (Chapter 4.3) was initially motivated by
keeping the number of features low, to reduce overfitting. This is not really needed
for linear SVMs, since they handle correlated features well. Furthermore the cal-
culation for inter-feature correlation gets really slow when the time and frequency
resolution is high. Due to this, the denominator containing the inter-feature correla-
tion term r̄ f f of equation (4.7) was ignored when a linear kernel was used.

The data was z-scored and then discretized. Features were then chosen via
correlation-based feature selection and fed into a SVM.

Genetic Algorithm
See Chapter 4.4 for an introduction to genetic algorithms. The data was first z-
scored. Then feature selection was done using a genetic algorithm. The fitness

65
Chapter 6. Method

function was the classifier performance for the target SVM. To avoid overfitting,
the features in the test and training set was randomized.

3. Dynamic Classifier
The idea to use a dynamic classifier came along late into the project. However, it
was still decided to introduce the concept and make a quick experiment.

Features
Feature definition In this work a feature was a time series containing the mag-
nitude of a frequency over a channel. It is possible to let each time series have its
own start and finish time. But to keep it simple for this initial experiment each time
series was forced to start at time zero and end at time T . Time T was chosen by
inspection in Figure 6.2 that shows where features that passed the ANOVA test on
an α = 0.1 level.

So the feature space consisted of all combinations of frequencies and channels. One
example would be the time series for the frequency 5 Hz at channel ten.
Feature Selection For feature selection ANOVA tests were calculated for all time
points, channels and frequencies. Then a time series was kept if q percent of the
samples had significant level above α.

Experiment
For classification, pairwise distance function (introduced in Chapter 3.5) was used.
Dynamic time warping was used as the distance metric with the distance as cost
functions. All features were forced to have the same warping. The total distance
was calculated using the L 1 norm over all channels.

The pairwise distance features were then fed into a linear SVM.

The result was compared with the ANOVA method choosing features from the same
feature set and another pair wise distance classifer that used the L 2 norm instead
of dynamic time warping for each channel, but the L 1 over the channels.

The goal was to see if this dynamic classifier could outperform the static and if the
dynamic time warping was useful.

66
 3. Dynamic Classifier

30

25

20
Channel

15

10

60
40
40 30
20 20
10
timepoint freq

Figure 6.2 Features that passed the ANOVA test on an α = 0.1 level. It can be seen that
there are more selected features for lower frequencies and earlier time points. Colors corre-
spond to channels.

67
7
Results

For all cases, 40 samples from each category was used for training. The remaining
trials was then used for testing. For subject 9 that was 24 faces, 22 objects and
22 landmarks. One subject was chosen to reduce work.This particular subject was
chosen since it gave the highest classification accuracy.

Frequency calculation was done using the function ft_freqanalysis in Field Trip
toolbox using (Morlet) wavelet and cfg.width = 5. The rest of the settings were
the defaults.

1. Static Classifier
Frequency is used from 5 to 49 Hz and time from 0 to 1 second. The full time
segment was used as the classifier should be able to handle if the features are worse
after a certain time point. The choice of 49 Hz was used since there is a lot of
disturbances of 50 Hz due to the power line. The lower limit was chosen as 5 Hz,
since low frequencies have bad time resolution and would lead to high delays if
used in a real-time BCI systems.

ANOVA
The ANOVA method was initially tested for other kernels than the linear. Those
results were however around chance and are not presented. The results for a linear
kernel are given in Table 7.1.

It was found that the best performance was achieved for a significance level of
α = 0.01.

Correlation-based feature selection

Correlation-based feature selection was tested for two cases. For the linear ker-
nel only class-feature correlation was used. For the non-linear kernels the feature-
feature correlation was also used. Uniform discretization of the features was used

68
 1. Static Classifier

Significance α time resolution number of features C accuracy %

0.05 0.1 1.8k 1 55.9
0.05 0.1/7 11k 1 67.7
0.01 0.1/7 5k 1 64.7
0.001 0.1/7 2.5k 1 61.8
Table 7.1 Results for ANOVA method for different configurations. Linear kernel was used.
The significance describes the significance for the one-way ANOVA test used during feature
selection. C is the regularization parameter for the SVM. Frequency resolution of 1 Hz was
used for all configuration.

for all experiments. The time resolution of 0.1 s and frequency resolution of 5 Hz
was used as that gave the best result.
Linear Kernel The SVM regularization parameter C , see (3.28), was set to 1, see
Table 7.2.

# features accuracy %
75 < 33
100 47.1
125 39.7
200 36.8
Table 7.2 Results for correlation based feature selection and linear SVM.

Radial basis function kernel The results were very bad for most parameter con-
figuration, the result presented in Table 7.3 is one which is acceptable. γ is the
parameter in the radial basis function, see (3.25).

# features C γ accuracy %
100 1 0.01 41.2
Table 7.3 Results for correlation based feature selection and radial basis function SVM.

None of the kernels could outperform with the ANOVA method.

Genetic Classifier
All genetic algorithm experiments were run with the following base configuration:

Crossover proability 0.8

elitism turned off
mutation rate 0.05
genome size 100

69
Chapter 7. Results

The fitness function was the classifier performance for the target classifier with
randomized training and test set for each generation. 70 percent of the available
samples were used to train the classifier for the fitness function and the remaining
30 percent were used for calculating the accuracy.

The best gene was always saved.

All accuracies are the mean of three runs, given in the last column

Linear Kernel For the linear kernel time resolution 0.1/7 s and frequency resolu-
tion 1 Hz was used. The regularization parameter C was 1 and the algorithm went
through 50 generations before termination, see Table 7.4.

number of features mean accuracy % accuracies %

500 60.3 58.8, 60.3 61.8
1000 59.8 55.9, 61.8, 61.8
2000 57.4 55.9, 55.9, 60.3
Table 7.4 Results for genetic algorithm selection and linear SVM.

Polynomial Kernel For the polynomial kernel the time resolution was 0.1 s and
frequency resolution was 5 Hz. No configuration of parameters gave results where
the accuracy for all the classes was above chance.

Radial basis function For the radial basis kernel the time resolution was after
some experimination chosen to be 0.05 s and frequency resolution was 4 Hz. The
algorithm went through 100 generations before termination, see Table 7.5.

number of features C γ mean accuracy % accuracies %

10 1 0.01 51.5 51.5, 50.0, 52.9
10 1 0.1 48.5 47.0, 45.6, 52.9
100 0.01 0.01 51.5 51.5, 51.5, 51.5
Table 7.5 Results for genetic algorithm feature selection and radial basis function SVM.

Summary
Here follows the best results for the different classifiers

70
 2. Dynamic Classifier

feature selection classifier accuracy %

ANOVA Linear SVM 67.7
Genetic algorithm Linear SVM 60.3
Genetic algorithm Radial basis SVM 51.5
CFS Linear SVM 47.1
CFS Radial basis SVM 41.2

None of the proposed method seems to be able to beat the ANOVA method.

2. Dynamic Classifier
For the dynamic classifier, time from 0 to 0.25 s with a time resolution of 1/3 s
was used. The frequency from 5 to 15 Hz was used with a resolution of 1 Hz. This
was chosen as it seemed like by visual inspection that configuration contained many
good time series. The time resolution was decided after some experiments.

The dynamic classifier kept all time series where all samples (q = 100%) were
significant on p = 0.2. Three different classifiers were trained. One based on the
ANOVA method with significance level p = 0.01. Two dynamic, one where the
distance function was the dynamic time warping, and one where the distance was
the L1 norm. The results can be seen in Table 7.6.

Method Total accuracy % Face acc % Landmark acc % Object acc %

ANOVA 45.6 79.2 36.4 18.2
DTW dynamic 41.2 62.6 31.8 27.3
L1 dynamic 38.2 58.3 31.8 22.7
Table 7.6 The results of the dynamic time warping experiment.

Here the accuracy for the categories are displayed as there is a large variation be-
tween them and some of them were below chance. The reason for this large variation
might be that features that were important to distinguish between landmarks and ob-
jects were discarded by the visual inspection that chose the features to use for this
experiment.

The dynamic classifier could not beat the ANOVA method, but it was not the most
well developed classifier. The dynamic time warping proved to slightly increase the
classifier results compared to the standard L1 norm.

71
8
Discussion and
Conclusions

1. Conclusions
The BCI technique is promising and is able to achieve high accuracy. However,
so far no system manage to be both fast and accurate without having to focus on
a stimuli. In addition, the gear is very immobile and sensitive. For industrial use,
the speed of the systems must be increased, while atleast maintaining accuracy. For
industrial use, the sensors needs to achieve a higher combination of portability and
quality.

The results of the classifier on the data used could not be improved by the two
feature selection method tested.

The dynamic time warping experiment showed that there is some information to be
gained by using the time dynamics, however the magnitude of that information was
very low for this particular dynamic classifier.

2. Discussion
For the experimental part of this work two things can be noted. Firstly, the genetic
algorithm had its best accuracy for fewer features than the ANOVA method for
linear kernel. That might be an indication that there is some overfitting problem
in the configuration and implementation of the genetic algorithm. Even with that
solved, it seems far fetched that it could beat the ANOVA method.

Secondly, the results for the non linear kernels were worse than the linear ones.
Furthermore they had to use small γ so they were not that complex. It seems like it
is best to use a linear kernel for the data set and the features used.

72
 3. Future Work

3. Future Work
There are some techniques that could be explored in future work to improve the
classifier. Outlier rejection or sample weighting, where the error penalty depends
on each samples could help deal with the possible noisy samples. Feature weight-
ing could also be explored. At the moment all features are assumed to be equally
important as they are normalized the same way. Features could be normalized dif-
ferently depending on how good they seem to be. This would allow for different
features to have different impact.

It is the authors opinion that the dynamic classifier should be explored further. The
first thing to implement would be to allow each time series to start and end at dif-
ferent times. Normalization of both time series and pair wise distance could be ex-
plored. Another dynamic classifier could also be used, for example a hidden Markov
model. Pair wise distance function SVM was used for its simplicity. Even if it is not
superior on this data set it is expected to perform better in non-clinical applications
were the signals are not as well aligned.

Features and classifier should be chosen with the problem in mind. But this work
used simple features and a classifier that is known to do well with high dimensional
data. Overall the approach have been of a somewhat black box nature. We have
some inputs coming from something and we want to classify them.

This was fine for this project as it tried to improve upon old results and used the
same setup as previous work did. But it is my opinion that combining knowledge of
different field to produce more meaningful features is likely to improve the result. A
more advance feature could, for example, take into account interactions between dif-
ferent parts of the brain explicitly. This will require cooperation between researches
from different fields.

I also think inspiration could be gained from the signal processing field were dy-
namic classifiers have been studied for a long time. I envision that most BCI classi-
fiers will be dynamic just as the process is.

73
Bibliography

Ang, K. K., C. Guan, K. S. G. Chua, B. T. Ang, C. W. K. Kuah, C. Wang, K. S. Phua,

Z. Y. Chin, and H. Zhang (2008). “A clinical evaluation of non-invasive motor
imagery-based brain-computer interface in stroke”. In: 2008 30th Annual Inter-
national Conference of the IEEE Engineering in Medicine and Biology Society.
IEEE, pp. 4178–4181.
Berg, A. T., S. F. Berkovic, M. J. Brodie, J. Buchhalter, J. H. Cross, W. Van Emde
Boas, J. Engel, J. French, T. A. Glauser, G. W. Mathern, S. L. Moshé, D. Nordli,
P. Plouin, and I. E. Scheffer (2010). “Revised terminology and concepts for or-
ganization of seizures and epilepsies: report of the ilae commission on clas-
sification and terminology, 2005–2009”. Epilepsia 51:4, pp. 676–685. ISSN:
1528-1167. DOI: 10 . 1111 / j . 1528 - 1167 . 2010 . 02522 . x. URL: http :
//dx.doi.org/10.1111/j.1528-1167.2010.02522.x.
Blankertz, B., G. Dornhege, M. Krauledat, K.-R. Muller, V. Kunzmann, F. Losch,
and G. Curio (2006). “The Berlin Brain-Computer Interface: EEG-based com-
munication without subject training”. IEEE transactions on neural systems and
rehabilitation engineering 14:2, pp. 147–152.
Brémaud, P. (2013). Markov chains: Gibbs fields, Monte Carlo simulation, and
queues. Vol. 31. Springer Science & Business Media, New York.
Brunner, C., R. Scherer, B. Graimann, G. Supp, and G. Pfurtscheller (2006). “On-
line control of a brain-computer interface using phase synchronization”. IEEE
Transactions on Biomedical Engineering 53:12, pp. 2501–2506.
Chae, Y., J. Jeong, and S. Jo (2012). “Toward brain-actuated humanoid robots: asyn-
chronous direct control using an eeg-based bci”. IEEE Transactions on Robotics
28:5, pp. 1131–1144.
Chang, C.-C. and C.-J. Lin (2011). “Libsvm: a library for support vector machines”.
ACM Transactions on Intelligent Systems and Technology (TIST) 2:3. Software
available at http : / / www . csie . ntu . edu . tw / ~cjlin / libsvm, accessed
2016-07-15, p. 27.

74
 Bibliography

Chong, M. S. T. (2013). Parameter and state estimation of nonlinear systems with

applications in neuroscience. PhD thesis. University of Melbourne, Department
of Electrical and Electronic Engineering.
Cortes, C. and V. Vapnik (1995). “Support-vector networks”. Machine Learning
20:3, pp. 273–297. DOI: 10 . 1023 / A : 1022627411411. URL: http : / / dx .
doi.org/10.1023/A:1022627411411.
Cristianini, N. and J. Shawe-Taylor (2000). An introduction to support vector ma-
chines and other kernel-based learning methods. Cambridge university press,
Cambridge.
Cuturi, M. (2011). “Fast global alignment kernels”. In: Proceedings of the 28th
international conference on machine learning (ICML-11), pp. 929–936.
D. Gupta, S. G. (2012). “An overview of methods maintaining diversity in genetic
algorithms”. International Journal of Emergin Technology and Avanced Engi-
neering 2:5, pp. 56–60.
Daly, I., S. J. Nasuto, and K. Warwick (2011). “Single tap identification for fast BCI
control”. Cognitive Neurodynamics 5:1, pp. 21–30. ISSN: 1871-4099.
Daly, I., S. J. Nasuto, and K. Warwick (2012). “Brain computer interface control via
functional connectivity dynamics”. Pattern recognition 45:6, pp. 2123–2136.
Duda, R. O., P. E. Hart, and D. G. Stork (2000). Pattern Classification (2Nd Edi-
tion). Wiley-Interscience, New York. ISBN: 0471056693.
“Dynamic Time Warping” (2007). In: Information Retrieval for Music and Motion.
Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 69–84. ISBN: 978-3-540-
74048-3. DOI: 10.1007/978- 3- 540- 74048- 3_4. URL: http://dx.doi.
org/10.1007/978-3-540-74048-3_4.
Eiben, A. E., R. Hinterding, and Z. Michalewicz (1999). “Parameter control in evo-
lutionary algorithms”. IEEE Transactions on Evolutionary Computation 3:2,
pp. 124–141. ISSN: 1089-778X. DOI: 10.1109/4235.771166.
Fayyad U & Irani, K (1993). “Multi-interval discretization of contrinuous-values
attributes for classification learning”. In: Proceedings of the Thirteenth Interna-
tional Joint Conference on Artifical Intelligence, pp. 1022–1027.
Fazel-Rezai, R., B. Z. Allison, C. Guger, E. W. Sellers, S. C. Kleih, and A.
Kübler (2012). “P300 brain computer interface: current challenges and emerg-
ing trends”. Frontiers in Neuroengineering 5:14. ISSN: 1662-6443. DOI: 10 .
3389 / fneng . 2012 . 00014. URL: http : / / www . frontiersin . org /
neuroengineering/10.3389/fneng.2012.00014/abstract.
Feyereisen, T. The pilot brain. accessed online https://aerospace.honeywell.
com/en/blogs/2016/april/the-pilot-brain 2016-06-30.

75
Bibliography

Gudmundsson, S., T. P. Runarsson, and S. Sigurdsson (2008). “Support vector ma-

chines and dynamic time warping for time series”. In: 2008 IEEE International
Joint Conference on Neural Networks (IEEE World Congress on Computational
Intelligence), pp. 2772–2776. DOI: 10.1109/IJCNN.2008.4634188.
Hall, M. A. (1998). Correlation-based feature selection for machine learning. PhD
thesis. University of Waikato, Department of Computer Science.
Hawkins, D. M. (2004). “The problem of overfitting”. Journal of chemical informa-
tion and computer sciences 44:1, pp. 1–12.
Hjorth, B. (1970). “EEG analysis based on time domain properties”. Electroen-
cephalography and Clinical Neurophysiology 29:3, pp. 306 –310. ISSN: 0013-
4694. DOI: http : / / dx . doi . org / 10 . 1016 / 0013 - 4694(70 ) 90143 -
4. URL: http : / / www . sciencedirect . com / science / article / pii /
0013469470901434.
Hramov, A. E., A. A. Koronovskii, V. A. Makarov, A. N. Pavlov, and E. Sitnikova
(2015). Wavelets in Neuroscience. Springer-Verlag, Berlin Heidelbergs.
Hsu, C.-W. and C.-J. Lin (2002). “A comparison of methods for multiclass support
vector machines”. IEEE Transactions on Neural Networks 13:2, pp. 415–425.
Huang, S. H. (2015). “Supervised feature selection: a tutorial”. Artifical Intelligence
Research 4:2, pp. 22–32.
Hwang, H.-J., S. Kim, S. Choi, and C.-H. Im (2013). “EEG-based brain-computer
interfaces: a thorough literature survey”. International Journal of Human-
Computer Interaction 29:12, pp. 814–826. DOI: 10.1080/10447318.2013.
780869. eprint: http://dx.doi.org/10.1080/10447318.2013.780869.
URL: http://dx.doi.org/10.1080/10447318.2013.780869.
Jafarpour, A., L. Fuentemilla, A. J. Horner, W. Penny, and E. Duzel (2014). “Re-
play of very early encoding representations during recollection”. The Journal of
Neuroscience 34:1, pp. 242–248.
Jakobsson, A. (2013). An Introduction to Time Series Modeling. 2nd edition. Stu-
dentlitteratur, Lund.
Joachims, T. (1998). “Text categorization with support vector machines: learning
with many relevant features”. In: Nédellec, C. et al. (Eds.). Machine Learning:
ECML-98: 10th European Conference on Machine Learning Chemnitz, Ger-
many, April 21–23, 1998 Proceedings. Springer Berlin Heidelberg, Berlin, Hei-
delberg, pp. 137–142. ISBN: 978-3-540-69781-7. DOI: 10.1007/BFb0026683.
URL: http://dx.doi.org/10.1007/BFb0026683.
Juang, B. H. and L. R. Rabiner (1991). “Hidden Markov Models for Speech Recog-
nition”. Technometrics 33:3, pp. 251–272. DOI: 10.1080/00401706.1991.
10484833. eprint: http://www.tandfonline.com/doi/pdf/10.1080/
00401706.1991.10484833. URL: ttp://www.tandfonline.com/doi/
abs/10.1080/00401706.1991.10484833.

76
 Bibliography

Krusienski, D. J., D. J. McFarland, and J. R. Wolpaw (2006). “An evaluation of

autoregressive spectral estimation model order for brain-computer interface ap-
plications”. In: Engineering in Medicine and Biology Society, 2006. EMBS’06.
28th Annual International Conference of the IEEE. IEEE, pp. 1323–1326.
LaFleur, K., K. Cassady, A. Doud, K. Shades, E. Rogin, and B. He (2013).
“Quadcopter control in three-dimensional space using a noninvasive motor
imagery-based brain–computer interface”. Journal of Neural Engineering 10:4,
p. 046003. URL: http://stacks.iop.org/1741-2552/10/i=4/a=046003.
Lee, T.-W. (1998). “Independent component analysis”. In: Independent Component
Analysis: Theory and Applications. Springer US, Boston, MA, pp. 27–66. ISBN:
978-1-4757-2851-4. DOI: 10.1007/978- 1- 4757- 2851- 4_2. URL: http:
//dx.doi.org/10.1007/978-1-4757-2851-4_2.
Lei, H. and B. Sun (2007). “A study on the dynamic time warping in kernel ma-
chines”. In: Signal-Image Technologies and Internet-Based System, 2007. SITIS
’07. Third International IEEE Conference on, pp. 839–845. DOI: 10 . 1109 /
SITIS.2007.112.
Liu, Y. and T. Khoshgoftaar (2004). “Reducing overfitting in genetic programming
models for software quality classification”. In: High Assurance Systems Engi-
neering, 2004. Proceedings. Eighth IEEE International Symposium on, pp. 56–
65. DOI: 10.1109/HASE.2004.1281730.
Lotte, F, M Congedo, A Lécuyer, F Lamarche, and B Arnaldi (2007). “A review
of classification algorithms for EEG-based brain–computer interfaces”. Journal
of Neural Engineering 4:2, R1. URL: http : / / stacks . iop . org / 1741 -
2552/4/i=2/a=R01.
Murakami, S. and Y. Okada (2006). “Contributions of principal neocortical neurons
to magnetoencephalography and electroencephalography signals”. The Jour-
nal of Physiology 575:3, pp. 925–936. ISSN: 1469-7793. DOI: 10 . 1113 /
jphysiol.2006.105379. URL: http://dx.doi.org/10.1113/jphysiol.
2006.105379.
Nicolas-Alonso, L. F. and J. Gomez-Gil (2012). “Brain computer interfaces, a re-
view”. Sensors 12:2, pp. 1211–1279.
Obermaier, B, C Guger, C Neuper, and G Pfurtscheller (2001). “Hidden Markov
models for online classification of single trial EEG data”. Pattern Recognition
Letters 22:12. Selected Papers from the 11th Portuguese Conference on Pattern
Recognition - {RECPAD2000}, pp. 1299 –1309. ISSN: 0167-8655. DOI: http:
//dx.doi.org/10.1016/S0167-8655(01)00075-7. URL: http://www.
sciencedirect.com/science/article/pii/S0167865501000757.
Oostenveld, R., P. Fries, E. Maris, and J.-M. Schoffelen (2011). “Fieldtrip: open
source software for advanced analysis of MEG, EEG, and invasive electrophys-
iological data”. Intell. Neuroscience 2011, 1:1–1:9. ISSN: 1687-5265. DOI: 10.
1155/2011/156869. URL: http://dx.doi.org/10.1155/2011/156869.

77
Bibliography

Penny, W. D., K. J. Friston, J. T. Ashburner, S. J. Kiebel, and T. E. Nichols (2011).

Statistical parametric mapping: the analysis of functional brain images. Aca-
demic press, Great Britain.
Percival, D. B. and A. T. Walden (1993). Spectral analysis for physical applications.
Cambridge University Press.
Rejer, I. and K. Lorenz (2013). “Genetic algorithm and forward selection for feature
selection in eeg feature space”. Journal of Theoretical and Applied Computer
Science 7:2, pp. 72–82.
Sandsten, M. (2016). Time-frequency analysis of time-varying signals and non-
stationary processes. Lund University, Centre for Mathematical Sciences, avail-
able online at http://www.maths.lth.se/matstat/kurser/masm26/
2016/TIMEFREQkompendie.pdf.
Schaap, P. (2016). Solving paralysis using brain computer interfaces. http : / /
europe . newsweek . com / brain - computer - interface - technology -
469920?rm=eu accessed 2016-09-25.
Scheffé, H. (1959). The Analysis of Variance. John Wiley & Sons, Inc., New York.
Schlögl, A., K Lugger, and G Pfurtscheller (1997). “Adaptive autoregressive param-
eters for a brain-computer-interface experiment”. In: Engineering in Medicine
and Biology Society. Vol. 4, pp. 1533–1535.
Srinivasan, R. (1999). “Methods to improve the spatial resolution of eeg”. Interna-
tional Journal of Bioelectromagnetism 1:1, pp. 102–111.
Thorpe, J., P. C. van Oorschot, and A. Somayaji (2005). “Pass-thoughts: authenti-
cating with our minds”. In: Proceedings of the 2005 Workshop on New Security
Paradigms. NSPW ’05. ACM, Lake Arrowhead, California, pp. 45–56. ISBN:
1-59593-317-4. DOI: 10.1145/1146269.1146282. URL: http://doi.acm.
org/10.1145/1146269.1146282.
Whitley, D. (1994). “A genetic algorithm tutorial”. Statistics and Computing 4:2,
pp. 65–85. ISSN: 1573-1375. DOI: 10.1007/BF00175354. URL: http://dx.
doi.org/10.1007/BF00175354.
Widmaier, E. P., H. Raff, and K. T. Strang (2014). Vander’s Human Physiology: The
mechanisms of Body Function. 13th ed. Mc Graw-Hill, New York.

78

/XQG8QLYHUVLW\ Document name
0$67(567+(6,6
'HSDUWPHQWRI$XWRPDWLF&RQWURO Date of issue
%R[ 2FWREHU
6(/XQG6ZHGHQ Document Number
,651/87)'7)576(
Author(s) Supervisor
0DUWLQ+H\GHQ -RKDQ(NHU(ULFVVRQ
0LNDHO-RKDQVVRQ'HSDUWPHQWRI3V\FKRORJ\/XQG
8QLYHUVLW\6ZHGHQ
%R%HUQKDUGVVRQ'HSWRI$XWRPDWLF&RQWURO/XQG
8QLYHUVLW\6ZHGHQ
$QGHUV5REHUWVVRQ'HSWRI$XWRPDWLF&RQWURO/XQG
8QLYHUVLW\6ZHGHQ H[DPLQHU 
Sponsoring organization

Title and subtitle

&ODVVLILFDWLRQRI((*GDWDXVLQJPDFKLQHOHDUQLQJWHFKQLTXHV

Abstract

$XWRPDWLFLQWHUSUHWDWLRQRIUHDGLQJIURPWKHEUDLQFRXOGDOORZIRUPDQ\LQWHUHVWLQJDSSOLFDWLRQV
LQFOXGLQJPRYHPHQWRISURVWKHWLFOLPEVDQGPRUHVHDPOHVVPDQPDFKLQHLQWHUDFWLRQ
7KLVZRUNVWXGLHGFODVVLILFDWLRQRI((*VLJQDOVXVHGLQDVWXG\RIPHPRU\7KHJRDOZDVWRHYDOXDWH
WKHSHUIRUPDQFHRIWKHVWDWHRIWKHDUWDOJRULWKPV$VHFRQGDU\JRDOZDVWRWU\WRLPSURYHXSRQWKH
UHVXOWRIDPHWKRGWKDWZDVXVHGLQDVWXG\VLPLODUWRWKHRQHXVHGLQWKLVZRUN
)RUWKHH[SHULPHQWWKHVLJQDOVZHUHWUDQVIRUPHGLQWRWKHIUHTXHQF\GRPDLQDQGWKHLUPDJQLWXGHV
ZHUHXVHGDVIHDWXUHV$VXEVHWRIWKHVHIHDWXUHVZDVWKHQVHOHFWHGDQGIHGLQWRDVXSSRUWYHFWRU
PDFKLQHFODVVLILHU7KHILUVWSDUWRIWKLVZRUNWULHGWRLPSURYHWKHVHOHFWLRQRIIHDWXUHVWKDWZDVXVHG
WRGLVFULPLQDWHEHWZHHQGLIIHUHQWPHPRU\FDWHJRULHV7KHVHFRQGSDUWLQYHVWLJDWHGWKHXVHVRIWLPH
VHULHVDVIHDWXUHVLQVWHDGRIWLPHSRLQWV
7ZRIHDWXUHVHOHFWLRQPHWKRGVJHQHWLFDOJRULWKPDQGFRUUHODWLRQEDVHGZHUHLPSOHPHQWHGDQG
WHVWHG%RWKRIWKHPSHUIRUPHGZRUVHWKDQWKHEDVHOLQH$129$PHWKRG
7KHWLPHVHULHVFODVVLILHUDOVRSHUIRUPHGZRUVHWKDQWKHVWDQGDUGFODVVLILHU+RZHYHUH[SHULPHQWV
VKRZHGWKDWWKHUHZDVLQIRUPDWLRQWRJDLQE\XVLQJWKHWLPHVHULHVPRWLYDWLQJPRUHDGYDQFHG
PHWKRGVWREHH[SORUHG
%RWKWKHUHVXOWVDFKLHYHGE\WKLVWKHVLVDQGLQRWKHUZRUNDUHDERYHFKDQFH+RZHYHUKLJK
DFFXUDFLHVFDQRQO\EHDFKLHYHGDWWKHFRVWRIORQJGHOD\VDQGIHZRXWSXWDOWHUQDWLYHV7KLVOLPLWVWKH
LQIRUPDWLRQWKDWFDQEHH[WUDFWHGIURPWKH((*VHQVRUDQGLWVXVDELOLW\

Keywords


Classification system and/or index terms (if any)



Supplementary bibliographical information


ISSN and key title ISBN
 
Language Number of pages Recipient’s notes
(QJOLVK 
Security classification

KWWSZZZFRQWUROOWKVHSXEOLFDWLRQV