J. Biomedical Science and Engineering, 2010, 3, 612-617
doi:10.4236/jbise.2010.36083 Published Online June 2010 (http://www.SciRP.org/journal/jbise/
JBiSE
).
Published Online June 2010 in SciRes. http://www.scirp.org/journal/jbise
Performance comparison of neural network training methods
based on wavelet packet transform for classification of five
mental tasks
Vijay Khare1, Jayashree Santhosh2, Sneh Anand3, Manvir Bhatia4
1Jaypee Institute of Information Technology, Department of Electronics and communication Engineering, New Delhi, India;
2Indian Institute of Technology, Computer Services Centre, New Delhi, India;
3Indian Institutes of Technology, Centre for Biomedical Engineering, New Delhi, India;
4Department of Sleep Medicine, New Delhi, India.
Received 11 February 2010; revised 15March 2010; accepted 25 March 2010.
ABSTRACT
In this study, performances comparison to discrimi-
nate five mental states of five artificial neural net-
work (ANN) training methods were investigated.
Wavelet Packet Transform (WPT) was used for fea-
ture extraction of the relevant frequency bands from
raw electroencephalogram (EEG) signals. The five
ANN training methods used were (a) Gradient De-
scent Back Propagation (b) Levenberg-Marquardt (c)
Resilient Back Propagation (d) Conjugate Learning
Gradient Back Propagation and (e) Gradient Descent
Back Propagation with movementum.
Keywords: Electroencephalogram (EEG); Wavelet
Packet Transform (WPT); Artificial Neural Network
(ANN)
1. INTRODUCTION
Brain computer interface use a non muscular communi-
cation channel for conveying message and command to
the external words in the absence of biological channels
[1-3]. Neuromuscular disorders like Amyotrophic lateral
sclerosis can temporarily or permanently impair spoken
and physical communication. Those most severely af-
fected may lose all voluntary muscle control and may be
completely locked in to their bodies, unable to commu-
nicate in any way. Using cognitive abilities is sometimes
the only the way to restore communication and motor
function [4-7]. Through training, subjects can learn to
control their brain activity in predefined fashion that is
classified by pattern recognition algorithms. Accuracy of
classification is affected by the quality of EEG signals
and the processing algorithms. The processing algo-
rithms include preprocessing, feature extraction and
classification. Previous studies investigated the effect of
different feature extraction algorithms with along and
the different mental tasks on classification accuracy was
investigated [8,9].
In this study, wavelet packet transform (WPT) method
was used to capture the information of mental tasks from
eight channel EEG signals of nine subjects. The coeffi-
cients of wavelet packet transform (WPT) were used as
the best fitting input vector for ANN. Five various artifi-
cial neural networks (ANN) training methods were used
to compare the performance in discrimination of five
mental tasks.
2. METHODOLOGY
2.1. Subjects
Nine right-handed healthy male subjects of age (mean
23yr) having no sign of any motor- neuron diseases were
selected for the study. A pro-forma was filled in with
detail of their age & education level. The participants
were student volunteers for their availability and interest
in the study. EEG data was collected after taking written
consent for participation. Full explanation of the ex-
periment was provided to each of the participants.
2.2. EEG Data Acquisition
EEG Data used in this study was recorded on a Grass
Telefactor EEG Twin3 Machine available at Deptt. of
Neurology, Sir Ganga Ram Hospital, New Delhi. EEG
recording was done for five mental tasks for five days,
from nine selected subjects. Data was recorded for 10
sec during each task and each task was repeated five
times per session per day. Bipolar and Referential EEG
was recorded using eight standard positions C3, C4, P3,
P4, O1 O2, and F3, F4 by placing gold electrodes on
scalp, as per the international 10-20 standard system of
electrode placement as shown in Figure 1. The settings

V. Khare et al. / J. Biomedical Science and Engineering 3 (2010) 612-617 613
Figure 1. Montage for present study.
Relax Mental
Task
Sec
…
0
20 10
Figure 2. Timing of the Protocol.
used for data collection were: low pass filter 1Hz, high
pass filter 35 Hz, sensitivity 150 micro volts/mm and
sampling frequency fixed at 400 Hz. The reference elec-
trodes were placed on ear lobes and ground electrode on
the forehead. EOG (Electooculargram) being a noise
artifact, was derived from two electrodes placed on outer
canthus of left and right eye in order to detect and elimi-
nate eye movement artifacts.
2.3. Experiment Paradigm
An experiment paradigm was designed for the study and
the protocol was explained to each participant before the
experiment. In this, the subject was asked to comfortably
lie down in a relaxed position with eyes closed. After
assuring the normal relaxed state by checking the status
of alpha waves, the EEG was recorded for 50 sec, col-
lecting five session of 10 sec epoch for the relaxed state.
This was used as the baseline reference for further analy-
sis of mental task. The subject was asked to perform a
mental task on presentation of an audio cue. Five session
of 10 sec epoch for each mental task were recorded as
shown in Figure 2. The whole experimental lasted for
about one hour including electrodes placement.
Data collected from nine subjects performing five
mental tasks were analyzed. The following mental tasks
were used:
Relaxed: The subject was asked to relax with their
eyes closed. No mental or physical task to be performed
at this stage.
Arithmetic Task: The subject was asked to perform
simple arithmetic (SA) and complex arithmetic (CA). An
example of a trivial calculation is to multiply 2 by 3 and
nontrivial task is to multiply 49 by 78. The subject was
instructed not to vocalize or make movements while
solving the problem. EEG signal were recorded corre-
sponding.
Geometric Figure Rotation (R): The subject was given
30 seconds to see complex three dimensional objects,
after which the object was removed. The subject was
instructed to visualize the object being rotated about an
axis. The EEG signals were recorded during this period.
Movement Imagery (M): The subject was asked to
plan movement of the right hand and corresponding
EEG signals were recorded during this period.
2.4. Feature Extraction
The frequency spectrum of the signal was first analyzed
through Fast Fourier Transform (FFT) method [10]. The
FFT plot of signals from the all electrode pairs were ob-
served and maximum average change in EEG amplitude
was noted as shown in Figures 3-7.
For relaxed, the peaks of power spectrum almost co-
incide for central area in the alpha frequency range (8-13
Hz)[11]. EEG recorded with relaxed state is considered
to be the base line for the subsequent analysis. Mu
rhythms are generated over sensorimotor cortex during
planning a movement. For movement imagery (M) of
right hand, maximum up to 50% band power attenuation
was observed in contralateral (C3 w.r.t C4) hemisphere
in the alpha frequency range (8-13 Hz) [11,12]. For geo-
metrical figure rotation(R), the peak of the power spec-
trum was increased in right hemisphere rather than left
in the occipital area for the alpha frequency range (8-13
Hz) [13]. For simple arithmetic(SA), the peak of the
power spectrum was increased in left hemisphere rather
than right hemisphere in the frontal area for the alpha
frequency range (8-13 Hz) [14].For complex arithmetic
(CA), the peak of the power spectrum was increased in
left hemisphere rather than right hemisphere in the parie-
tal area for the alpha frequency range (8-13 Hz).
0
50
100
150
200
C3 C4
Ele ct r ode
Power Spectrum Density for Alpha Frequency
Range
Figure 3. Power Spectra for Relax state at C3 and C4 channel.

614 V. Khare et al. / J. Biomedical Science and Engineering 3 (2010) 612-617
F4
F3
S1
0
50
100
150
Electrode
Power Spectrum Density for Alpha Frequency
Range
Figure 4. Power Spectra for simple arithmetic at F3 and F4 channel.
0
50
100
150
200
O2 O1
Electr ode
Power Spectrum Density for Alpha Frequency
Range
Figure 5.Power Spectra for rotation at O1 and O2 channel.
0
50
100
150
200
250
300
P4 P3
Electrode
Range
Power Spectrum Density for Alpha Frequency
Figure 6. Power Spectra for complex Arithmetic P3 and P4 channel.
0
50
100
150
200
C3 C4
El e ctr od e
Power Spectrum Density for Alpha Frequency
Range
Figure 7. Power Spectra for right movement at C3 and C4 channel.
2.5. Wavelet Packet Transform
By applying Wavelet packet analysis on the original
signal wavelet coefficients in the 8-13 HZ frequency
band at the 5th level node (5, 3) were obtained. The
signal was reconstructed at node (5, 3). These coeffi-
cients are scaled and WPT coefficients are used as the
best fitting input vector for ANN. Wavelet transform
we were able to reduce 1 second of EEG data to 21
coefficients for each mental tasks [15,16].
2.6. Classification
The main advantage of choosing artificial neural net-
work for classification was due to fact that ANN’s could
be used to solve problems, where description for the data
is not computable. ANN could be trained using data to
discriminate the feature. The five different training
methods used for Classification in the present study were
Gradient Descent method Resilient Back propagation,
Levenberg-Marquardt, Conjugate Gradient Descent and
Gradient Descent back propagation with movementum.
For classification a two layer neural networks was
used for the instance a topology of {10, 1} indicate a 21
input, 10 neurons in hidden layer and one output layer.
The neural network was designed to accept a 21 element
input vector and give a single output. One second seg-
ments of EEG were classified. The training set was
composed of 60% EEG trial per mental tasks and test set
was composed 40% EEG trial per mental task. The out-
put was designed to give 0 for baseline and 1 for task.
The neural network was trained for a fixed number of
epochs and the training is done using a five learning
techniques [17]. Parameter used for five training meth-
ods of neural network for classification of five mental
tasks as shown in the Table 1.
2.7. Evaluations
Performance (RC) is defined as ratio between correctly
classified patterns in the test set to the total number of
patterns in the test set in percentage.
Number of correctly classified test patterns
Rc Total number of patterns in the test set 
With the help of above formula we calculate the per-
formance of each method for each task [18].
3. RESULTS
For the classification of five mental tasks neural network
training methods were used. Nine male right-handed
subjects participated in the experiments. The subjects
asked to perform five mental tasks namely relaxed, arith-
metic task (simple and complex multiplication), geomet-
rical figure rotation, and movement imagery. Table 2
showed the comparison of the performance of five neu-

V. Khare et al. / J. Biomedical Science and Engineering 3 (2010) 612-617 615
Table1. Parameter used for different back propagation algo-
rithms with topology {10, 1}.
Gradient Descent with Momentum (GDM)
Topology {10,1} α = 0.01. Mu = 0.01
MSE = 1e-5
Epoach = 5000
Gradient Descent method (GD)
Topology {10,1} α = 0.01
MSE = 1exp-(5)
Epoach = 5000
Resilient Back propagation (RBP)
Topology {10,1} α = 0.01
MSE = 1exp-(5)
Epoach = 5000 β = 0.75 and β1 = 1.05
Conjugate Gradient descent (CG)
Topology {10,1} α = 0.01
MSE = 1exp-(5)
Epoach = 5000
Levenberg-Marquardt (LM)
Topology {10,1} Mu = 0.01
MSE = 1exp-(5)
Epoach = 5000 Mu_dec = 0.1 and Mu_inc = 10
ral network (NN) training methods in classifying of five
mental tasks. From this table we can say that Resilient
Back Propagation method was most suitable for the clas-
sification of all five mental tasks. Because this method
gave highest performance (95%) all five mental tasks.
Percentage average accuracy shown in the Figures 8-
12.
4. CONCLUSIONS
In this study, nine healthy male subjects were se-
lected to investigate MLP classifier with various training
methods to discriminate five mental tasks (relaxed state,
movement imagery of right hand, geometrical figure
rotation, arithmetic simple task, and arithmetic complex
task) effectively. For relaxed, the peaks of power spec-
Gradient Descent back Propagation
80
85
90
95
100
SA CARM
mental tasks
pecentage of
accuracy
Figure 8. Classification accuracy using GDA BP train-
ing methods.
trum almost coincide in central area at a particular based
frequency. For simple arithmetic, it was observed that
the amplitude of the power spectrum for alpha frequency
range (8-13 Hz) increased left hemisphere rather than
right hemisphere in frontal region. For complex arithe-
matic, it was observed that the amplitude of the power
spectrum for alpha frequency range (8-13 Hz) increased
Resilient Back propagation
93
94
95
96
97
98
SA CARM
Mental tasks
percentage of
accuracy
Figure 9. Classification accuracy RBP training methods.
Conjugate Gradient BP
90
92
94
96
98
SA CARM
Mental tasks
percentage of
accu racy
Figure 10. Classification accuracy CGBP training methods.
Gradient descent method with
movementum
86
88
90
92
94
96
SA CARM
Mental tasks
percentage of
accuracy
Figure 11. Classification accuracy GDM training methods.
Levenberg-Marquardt
86
88
90
92
94
96
SA CARM
Mental tasks
percenatge of
accu racy
Figure 12. Classification accuracy LM training methods.

616 V. Khare et al. / J. Biomedical Science and Engineering 3 (2010) 612-617
Table 2. Comparisons of Different NN training Methods.
Tasks Baseline and
Arithmetic simple
Baseline and Arithmetic
complex Baseline and Rotation Baseline and movement
Techniques Correct
classification
Wrong
classifica-
tion
Correct
classifica-
tion
Wrong
classifica-
tion
Correct
classifica-
tion
Wrong
classifi-
cation
Correct
classifica-
tion
Wrong
classifica-
tion
Gradient Descent Back
Propagation 95% 5% 95% 5% 87.5% 12.5% 90% 10%
Resilient Back Propagation 97.5% 2. 5% 95% 5% 95% 5% 95% 5%
Conjugated Gradient BP 97.5% 2.5% 92.5% 7.5% 92.5% 7.5% 92.5% 7.5%
GD BP with Momentum 95% 5% 95% 5% 90% 10% 90% 10%
Levenberg-Marquardt 95% 5% 90% 10% 90% 10% 92.5% 7.5%
left hemisphere rather than right hemisphere in parietal
region. For geometrical figure rotation, the peak of the
power spectrum in the alpha frequency range (8-13 Hz)
increased right occipital area. For movement imagery,
the peak of the power spectrum in the alpha frequency
range (8-13 Hz) had an attenuation central area. The
result showed the performance of neural network with
various training method, for classifying of mental tasks
w.r.t baseline. Resilient backpropagation training method
has best performance among all the training methods for
classification of mental tasks w.r.t baseline.
JBiSE
The authors would like to extend the work with se-
verely disabled people and to customize the device as
per individual response and requirements. This kind of
system can also be used in a variety of applications like–
Environment control units (ECU’S), helping disable
people to directly interact with hand held devices such as
cell phones and PDAs.
5. ACKNOWLEDGEMENTS
The authors would like to acknowledge their gratitude to the Centre of
Biomedical Engineering of IIT New Delhi and Electronics and com-
munication department of JIIT Noida. The authors also thank the sci-
entific and technical staff of EEG Laboratory of Sir Ganga Ram hospi-
tal, New Delhi for the help in carrying out the experiment.
REFERENCES
[1] Wolpaw, J.R., Birbaumer, N., Mc Farland, D.J., Plurtscheller,
G. and Vaughan, T.M. (2002) Brain computer interfaces for
communication and control. Clinical Neurophysiology,
113(6), 767-791.
[2] Pfurtschelle, G., Flotzinger, D. and Kalcher, J. (1993)
Brain computer interface－A new communication device
for handicapped people. Journal of Microcomputer Ap-
plications, 16, 293-299.
[3] Wolpaw, J.R., Vaughan, T.M. and Donchin, E. (1996)
EEG based communication prospects and problems.
IEEE Transactions on Rehabilitation Engineering, 4,
425-430.
[4] Lotte, F., Congedo, M., Lecuyer, A., Lamarche, F. and
Arnaldi, B. (2007) A review of classification algorithms
for EEG bases brain computer interface. Journal of
neural Engineering, 4, 1-13.
[5] Wolpaw, J.R., Leob, G.E., Allison, B.Z., Donchin, E. and
Turner, J.N. (2006) BCI Meeting 2005－Wokshop on
signals and rerecording methods. IEEE Transactions on
Neural Systems and Rehabilitation Engineering, 14, 138-
141.
[6] Anderson, C.W., Stolz, E.A. and Shamsunder, S. (1998)
Multivariable autoregressive model for classification of
spontaneous electroencephalogram during mental tasks.
IEEE Transactions on Biomedical Engineering, 45, 277-
278.
[7] Boostani, R., Graimann, B., Moradi, M.H. and Plurf-
scheller, G. (2007) Comparison approach toward finding
the best feature and classifier in cue BCI. Journal of
Medical and Biological Engineering and Computing, 45,
403-413.
[8] Palaniappan, R. (2006) Utilizing gamma band to improve
mental task based brain－Computer interface design.
IEEE Transactions on Neural and Rehabilitation Systems
Engineering, 14, 299-303.
[9] Keirn, Z.A. and Aunon, J.I. (1990) A new mode of
communication between man and his Surroundings. IEEE
Transactions on Biomedical Engineering, 37, 1209-1214.
[10] Palaniappan, R. (2005) Brain computer interface design
using band powers extracted during mental task. Procee-
ding of the 2nd International IEEE EMBS Conference on
Neural Engineering, Arlington, 321-324.
[11] Pfurtscheller, G., Neuper, C., Schlogl, A. and Lugger, K.
(1998) Separability of EEG signals recorded during right
and left motor imagery using adaptive auto regressive
parameters. IEEE Transactions on rehabilitation Engin-
eering, 6, 316-325.
[12] Santhosh, J., Bhatia, M., Sahu, S. and Anand, S. (2004)
Quantitative EEG analysis for assessment to plan a task
in ALS patients, a study of executive function (planning)
in ALS. Journal of Cognitive Brain Research, 22, 59-66.
[13] Nikolaev, A.R. and Anokhin, A.P. (1998) EEG frequency
ranges during reception and mental rotation of two and
three dimensional objects. Neuroscience and Bheaviour
physiology, 29, 203-223.
[14] Osaka, M. (1984) Peak alpha frequency of EEG during a
mental task: Task difficulty and hemisphere difference.
Journal of Psychophysiology, 21, 101-105.
[15] Ting, W., Zhenga, Y.G., Bang-huaa, Y. and Hong, S.
(2008) EEG feature extraction based on wavelet packet

V. Khare et al. / J. Biomedical Science and Engineering 3 (2010) 612-617 617
decom-position for brain computer interface. Measurement,
Elsevier Journal, 41, 618-625.
[16] Akay, M. (1995) Wavelet in biomedical engineering.
Journal of Annals of Biomedical Engineering, 23, 529-
530.
[17] Ravi, K.V.R. and Palaniappan, R. (2006) Neural Network
Classification of Late gamma band electroenceph-alogram
features, Soft Computing, 10, 163-169.
[18] Cheng, M., Gao, X., Gao, S. and Xu, D. (2002) Design
and implementation of a brain computer interface with
high transfer rates. IEEE Transactions on Biomedical
Engineering, 49, 1181-1186.