A Generic Approach for Identification of 
Event Related Brain Potentials via a 
Competitive Neural Network Structure 
Daniel H. Lange 
Department of Electrical Engineering 
Technion -IIT 
Hifa 32000 
Israel 
e-mail: lange@turbo.technion.ac.il 
Hava T. Siegelmann 
Department of Industrial Engineering 
Technion -IIT 
Hifa 32000 
Israel 
e-mail: iehavaie.technion.ac.il 
Hillel Pratt 
Evoked Potential Laboratory 
Technion- IIT 
Haifa 32000 
Israel 
e-mail: hilleltx.t echnion.ac.il 
Gideon F. Inbar 
Department of Electrical Engineering 
Technion -IIT 
Haifa 32000 
Israel 
e-mail: inbaree.technion.ac.il 
Abstract 
We present a novel generic approach to the problem of Event Related 
Potential identification and classification, based on a competitive Neu- 
ral Net architecture. The network weights converge to the embedded 
signal patterns, resulting in the formation of a matched filter bank. 
The network performance is analyzed via a simulation study, exploring 
identification robustness under low SNR conditions and compared to 
the expected performance from an information theoretic perspective. 
The classifier is applied to real event-related potential data recorded 
during a classic odd-ball type paradigm; for the first time, within- 
session variable signal patterns are automatically identified, dismiss- 
ing the strong and limiting requirement of a-priori stimulus-related 
selective grouping of the recorded data. 
902 D. H. Lange, H. T. Siegelmann, H. Pratt and G. F. Inbar 
1 
INTRODUCTION 
1.1 EVENT RELATED POTENTIALS 
Ever since Hans Berger's discovery that the electrical activity of the brain can be 
measured and recorded via surface electrodes mounted on the scalp, there has been 
major interest in the relationship between such recordings and brain function. The 
first recordings were concerned with the spontaneous electrical activity of the brain, 
appearing in the form of rhythmic voltage oscillations, which later received the term 
electroencephalo#ram or EEG. Subsequently, more recent research has concentrated on 
time-locked brain activity, related to specific events, external or internal to the subject. 
This time-locked activity, referred to also as Event Related Potentials (ERP's), is 
regarded as a manifestation of brain processes related to preparation for or in response 
to discrete events meaningful to the subject. 
The ongoing electrical activity of the brain, the EEG, is comprised of relatively slow 
fluctuations, in the range of 0.1 - 100 Hz, with magnitudes of 10 - 100 uV. ERP's 
are characterized by overlapping spectra with the EEG, but with significantly lower 
magnitudes of 0.1 - 10 uV. The unfavorable Signal to Noise Ratio (SNR) requires 
filtering of the raw signals to enable analysis of the time-locked signals. The common 
method used for this purpose is signal averaging, synchronized to repeated occurrences 
of a specific event. Averaging-based techniques assume a deterministic signal within 
the averaged session, and thus signal variability can not be modeled unless a-priori 
stimulus- or response-based categorization is available; it is the purpose of this paper to 
provide an alternative working method to enhance conventional averaging techniques, 
and thus facilitating identification and analysis of variable brain responses. 
1.2 COMPETITIVE LEARNING 
Competitive learning is a well-known branch of the general unsupervised learning 
theme. The elementary principles of competitive learning are (Rumelhart & Zipset, 
1985): (a) start with a set of units that are all the same except for some randomly 
distributed parameter which makes each of them respond slightly differently to a set 
of input patterns, (b) limit the strength of each unit, and (c) allow the units to com- 
pete in some way for the right to respond to a given subset of inputs. Applying these 
three principles yields a learning paradigm where individual units learn to specialize 
on sets of similar patterns and thus become feature detectors. Competitive learning 
is a mechanism well-suited for regularity detection (Haykin, 1994), where there is a 
population of input patterns each of which is presented with some probability. The 
detector is supposed to discover statistically salient features of the input population, 
without a-priori categorization into which the patterns are to be classified. Thus the 
detector needs to develop its own featural representation of the population of input 
patterns capturing its most salient features. 
1.3 PROBLEM STATEMENT 
The complicated, generally unknown relationships between the stimulus and its asso- 
ciated brain response, and the extremely low SNR of the brain responses which are 
practically masked by the background brain activity, make the choice of a self orga- 
nizing structure for post-stimulus epoch analysis most appropriate. The competitive 
network, having the property that its weights converge to the actual embedded sig- 
nal patterns while inherently averaging out the additive background EEG, is thus an 
evident choice. 
A Generic Approach for Identification of Event Related Brain Potentials 903 
2 THE COMPETITIVE NEURAL NETWORK 
2.1 THEORY 
The common architecture of a competitive learning system appears in Fig. 1. The 
system consists of a set of hierarchically layered neurons in which each layer is connected 
via excitatory connections with the following layer. Within a layer, the neurons are 
divided into sets of inhibitory clusters in which all neurons within a cluster inhibit 
all other neurons in the cluster, which results in a competition among the neurons to 
respond to the pattern appearing on the previous layer. 
Let toji denote the synaptic weight connecting input node i to neuron j. A neuron 
learns by shifting synaptic weights from its inactive to active input nodes. If a neuron 
does not respond to some input pattern, no learning occurs in that neuron. When a 
single neuron wins the competition, each of its input nodes gives up some proportion of 
its synaptic weight, which is distributed equally among the active input nodes, fulfilling: 
- toji= 1. According to the standard competitive learning rule, for a winning neuron 
to an input vector i, the change Atoji is defined by: Atoji -- (i - toil), where 
 is a learning rate coefficient. The effect of this rule is that the synaptic weights 
of a winning neuron are shifted towards the input pattern; thus assuming zero-mean 
additive background EEG, once converged, the network operates as a raaiehed filler 
bank classifier. 
2.2 MATCHED FILTERING 
Fom an information theoretic perspective, once the network has converged, our clas- 
siftcation problem coincides with the general detection problem of known signals in 
additive noise. For simplicity, we shall limit the discussion to the binary decision 
problem of a known signal in additive white Gaussian noise, expandable to the M-ary 
detection in colored noise (Van Trees, 1968). 
Adopting the common assumption of EEG and ERP additivity (Gevins, 1984), and 
distinct signal categories, the competitive NN weights inherently converge to the general 
signal patterns embedded within the background brain activity; therefore the converged 
network operates as a matched /ilter bank. Assuming the simplest binary decision 
problem, the received signal under one hypothesis consists of a completely known signal, 
Vs(l), representing the EP, corrupted by an additive zero-mean Gaussian noise to(t) 
with variance 2; the received signal under the other hypothesis consists of the noise 
to(t) alone. Thus: 
= o < < 
m: = + o < < T 
For convenience we assume that f s2(t)dt - 1, so that//7 represents the signal energy. 
The problem is to observe r(t) over the interval [0, T] and decide whether Ho or H 
is true. It can be shown that the matched tilter is the optimal detector, its impulse 
response being simply the signal reversed in time and shifted: 
k(r) ---- s(T- r) (1) 
Assuming that there is no a-priori knowledge of the probability of signal presence, the 
total probability of error depends only on the SNR and is given by (Van Trees, 1968): 
Fig. 2 presents the probability of true detection: (a) as a function of SNR, for minimized 
error probability, and (b) as a function of the probability of false detection. These 
904 D. H. Lange, H. T. Siegelmann, H. Pratt and G. E Inbar 
results are applicable to our detection problem assuming approximate Gaussian EEG 
characteristics (Gersch, 1970), or optimally by using a pre-whitening approach (Lange 
et. al., 1997). 
Layer 2 
Inhibitory Cluters 
INPUT PATFER N 
Figure 1: The architecture of a compet- 
itive learning structure: learning takes 
place in hierarchically layered units, 
presented as filled (active) and empty 
(inactive) dots. 
Probabtfity of True Detection with Minimum Error 
_o.8 
o;[ 
-40 
SNR In 
Detection1 performance: SNR = +20, +10, 0, -10 and -20 dB 
Pro. ability of False Detection 
Figure 2: Detection performance. Top: 
probability of detection as a function of 
the SNR. Bottom: detection character- 
istics. 
4O 
2.3 NETWORK TRAINING AND CONVERGENCE 
Our net includes a 300-node input layer and a competitive layer consisting of single- 
layered competing neurons. The network weights are initialized with random values 
and trained with the standard competitive learning rule, applied to the normalized 
input vectors: 
(3) 
The training is applied to the winning neuron of each epoch, while increasing the bias of 
the frequently winning neuron to gradually reduce its chance of winning consecutively 
(eliminating the dead neuron effect (Freeman & Skapura, 1992)). Symmetrically, its 
bias is reduced with the winnings of other neurons. 
In order to evaluate the network performance, we explore its convergence by analyzing 
the learning process via the continuously adapting weights: 
pj(n) = ( .. Aw]i ; j = 1,2,...,C (4) 
where C represents the pre-defined number of categories. We define a set of classifica- 
tion confidence coefficients of the converged network: 
rj = 1- .j(N) 
mxj{p(N)} (S) 
Assuming existence of a null category, in which the measurements include only back- 
grouna noise (X,G), max{p(V)} corresponds to the noise variance. Zhus the values 
of Fj, the confidence coefficients, ranging from 0 to I (random classification to com- 
pletely separated categories), indicate the reliability of classification, which breaks down 
with the fall of SNR. Finally, it should be noted that an explicit statistical evaluation 
of the network convergence properties can be found in (Lange, 1997). 
A Generic Approach for Identification of Event Related Brain Potentials 905 
2.4 SIMULATION STUDY 
A simulation study was carried out to assess the performance of the competitive net- 
work classification system. A moving average (MA) process of order 8 (selected ac- 
cording to Akaike's condition applied to ongoing EEG (Gersch, 1970)), driven by a 
deterministic realization of a Gaussian white noise series, simulated the ongoing back- 
ground activity (n). An average of 40 single-trials from a cognitive odd-ball type 
experiment (to be explained in the Experimental Study), was used as the signal s(n). 
Then, five 100-trial ensembles were synthesized, to study the classification performance 
under variable SNR conditions. A sample realization and its constituents, at an SNR 
of 0 dB, is shown in Fig. 3. The simulation included embedding the signal s(n) in 
the synthesized background activity (n) at five SNR levels (-20,-10,0,+10, and +20 
dB), and training the network with 750 sweeps (per SNR level). Fig. 4 shows the 
convergence patterns and classification confidences of the two neurons, where it can be 
seen that for SNR's lower than -10dB the classification confidence declines sharply. 
Tenplate, Nose.  a OdB gnal reallZlahOn 
10 = Conveeg Pattr n 
10 3 
Corer ger'l Pattn 
SNR=-10 dB 
Figure 3: A sample single realization 
(dotted) and its constituents (signal - 
solid, noise - dashed). SNR - 0 dB. 
Figure 4: Convergence 
classification confidence 
varying SNR levels. 
patterns and 
values for 
The classification results, tested on 100 input vectors, 50 of each category, for each 
SNR, are presented in the table below; due to the competitive scheme, Positives and 
False Negatives as well as Negatives and False Positives are complementary. These 
empirical results are in agreement with the analytical results presented in the above 
Matched Filtering secttog. 
Table 1: Classification Results 
I II Pos Neg I FP I FN I 
snr=+20dB 100% 100% 0% 0% 
snr-+10dB 100% 100% 0% 0% 
snr= 0dB 100% 100% 0% 0% 
snr=-10dB 88% 92% 8% 12% 
snr=-20dB 58% 54% 46% 42% 
3 EXPERIMENTAL STUDY 
3.1 MOTIVATION 
An important task in ERP research is to identify effects related to cognitive processes 
triggered by meaningful versus non-relevant stimuli. A common procedure to study 
these effects is the classic odd-ball paradigm, where the subject is exposed to a random 
906 D. H. Lange, H. T. Siegelmann, H. Pratt and G. E Inbar 
sequence of stimuli and is instructed to respond only to the task-relevant (Target) ones. 
Typically, the brain responses are extracted via selective averaging of the recorded data, 
ensembled according to the types of related stimuli. This method of analysis assumes 
that the brain responds equally to the members of each type of stimulus; however 
the validity of this assumption is unknown in this case where cognition itself is being 
studied. Using our proposed approach, a-prior{ grouping of the recorded data is not 
required, thus overcoming the above severe assumption on cognitive brain function. 
The results of applying our method are described below. 
3.2 EXPERIMENTAL PARADIGM 
Cognitive event-related potential data was acquired during an odd-ball type paradigm 
from Pz referenced to the mid-lower jaw, with a sample frequency of 250 Hz (Lange et. 
al., 1995). The subject was exposed to repeated visual stimuli, consisting of the digits 
'3' and '5', appearing on a PC screen. The subject was instructed to press a push- 
button upon the appearance of '5' - the Target stimulus, and ignore the appearances 
of the digit '3'. 
With odd-ball type paradigms, the Target stimulus is known to elicit a prominent 
positive component in the ongoing brain activity, related to the identification of a 
meaningful stimulus. This component has been labeled P300, indicating its polarity 
(positive) and timing of appearance (300 ms after stimulus presentation). The param- 
eters of the P300 component (latency and amplitude) are used by neurophysiologists to 
assess effects related to the relevance of stimulus and level of attention (Lange et. al., 
199S). 
3.3 IDENTIFICATION RESULTS 
The competitive network was trained with 80 input vectors, half of which were Target 
ERP's and the other half were Non Target. The network converged after approximately 
300 iterations (per neuron), yielding a reasonable confidence coefficient of 0.7. 
A sample of two single-trial post-stimulus sweeps, of the Target and Non-Target aver- 
aged ERP templates and of the NN identified signal categories, are presented in Fig. 
5. The convergence pattern is shown in Fig. 6. The automatic identification procedure 
has provided two signal categories, with almost perfect matches to the stimulus-related 
selective averaged signals. The obtained categorization confirnis the usage of averaging 
methods for this classic experiment, and thus presents an important result in itself. 
4 DISCUSSION AND CONCLUSION 
A generic system for identification and classification of single-trial ERP's was presented. 
The simulation study demonstrated the powerful capabilities of the competitive neural 
net in classifying the low amplitude signals embedded within the large background 
noise. The detection performance declined rapidly for SNR's lower than -10dB, which 
is in general agreement with the theoretical statistical results, where loss of significance 
in detection probability is evident for SNR's lower than -20dB. Empirically, high 
classification performance was maintained with SNR's of down to -10dB, yielding 
confidences in the order of 0.7 or higher. 
The experimental study presented an unsupervised identification and classification of 
the raw data into Target and Non-Target responses, dismissing the requirement of 
stimulus-related selective data grouping. The presented results indicate that the noisy 
brain responses may be identified and classified objectively in cases where relevance of 
A Generic Approach for Identification of Event Related Brain Potentials 907 
the stimuli is unknown or needs to be determined, e.g. in lie-detection scenarios (Lange 
& Inbar, 1996), and thus open new possibilities in ERP research. 
>= o >. 
Figure 5: Top row: sample raw Target 
and Non-Target sweeps. Middle row: 
Target and Non-Target ERP templates. 
Bottom row: the NN categorized pat- 
terns. 
SNR = -5 d,B 
Figure 6: Convergence pattern of the 
ERP categorization process; conver- 
gence is achieved after 300 iterations 
per neuron. 
References 
[1] Freeman J.A. and Skapura D.M. Neural Networks: Algorithms, Applications, and Pro- 
gramming Techniques: Addison-Wesley Publishing Company, USA, 1992. 
[2] Gersch W., "Spectral Analysis of EEG's by Autoregressive Decomposition of Time Se- 
ries," Math. Biosc., vol. 7, pp. 205-222, 1970. 
[3] Gevins A.S., "Analysis of the Electromagnetic Signals of the Human Brain: Milestones, 
Obstacles, and Goals," IEEE Trans. Biotaed. Eng., vol. BME-31, pp. 833-850, 1984. 
[4] Haykin S. Neural Networks: A Comprehensive Foundation. Macmillan College Publishing 
Company, Inc., USA, 1994. 
[5] Lange D. H. Modeling and Estimation of Transient, Evoked Brain Potentials. D.Sc. dis- 
sertation, Techion - Israel Institute of Technology, 1997. 
[6] Lange D.H. and Inbar G.F., "Brain Wave Based Polygraphy," Proceedings of the IEEE 
EMBS96 - the 18th Annual International Conference of the IEEE Engineering on 
Medicine and Biology Society, Amsterdam, October 1996. 
[7] Lange D.H., Pratt H. and Inbar G.F., "Modeling and Estimation of Single Evoked Brain 
Potential Components", IEEE. Trans. Biotaed. Eng., vol. BME-44, pp. 791-799, 1997. 
[8] Lange D.H., Pratt H., and Inbar G.F., "Segmented Matched Filtering of Single Event 
Related Evoked Potentials," IEEE. Trans. Biotaed. Eng., vol. BME-42, pp. 317-321, 
1995. 
[9] Rumelhart D.E. and Zipset D., "Feature Discovery by Competitive Learning," Cognitive 
Science, vol. 9, pp. 75-112, 1985. 
[10] Van Trees H.L. Detection, Estimation, and Modulation Theory: Part 1: John Wiley and 
Sons, Inc., USA, 1968. 
