(PDF) Combaz 2010 Mlsp

ERROR-RELATED POTENTIAL RECORDED BY EEG IN THE CONTEXT OF A P300 MIND SPELLER BRAIN-COMPUTER INTERFACE Adrien Combaz1, Nikolay Chumerin1, Nikolay V. Manyakov1, Arne Robben1, Johan A. K. Suykens2 and Marc M. Van Hulle1 1 K.U.Leuven, Laboratorium voor Neuro- en Psychofysiologie, Herestraat 49, B-3000 Leuven, Belgium 2 K.U.Leuven, ESAT-SCD, Kasteelpark Arenberg 10, BE-3001 Heverlee, Belgium ABSTRACT The Mind Speller® is a Brain-Computer Interface (BCI) which enables subjects to spell text on a computer screen by detecting P300 Event-Related Potentials in their electro- encephalograms (EEG). This BCI application is of particular interest for disabled patients who have lost all means of Fig. 1. User display for the P300 Mind-speller BCI. Left: verbal and motor communication. Error-related Potentials intensification of a column of the matrix display. Right: Feedback moment: the identified symbol is displayed on the screen. (ErrP) in the EEG are generated by the subject’s perception of an error. We report on the possibility of using this ErrP In invasive BCIs, a micro-electrode array is implanted for improving the performance of our Mind Speller®. We in the brain (mainly in the motor or premotor frontal areas or tested 6 subjects and recorded several typing sessions for into the parietal cortex [3]), while in non-invasive BCIs, each of them. Responses to correct and incorrect mostly electroencephalograms (EEGs) are recorded from performances of the BCI are recorded and compared. The the scalp. There are several types of EEG-based BCIs; for shape of the received ErrP is compared to other studies. The example some are based on Steady State Visually Evoked detection of this ErrP and its integration in the Mind Potential (SSVEP, [4]); they work by detecting the activity Speller® are discussed. of the brain at a specific frequency corresponding to the flickering frequency of a visual stimulus (see [5], [6] for 1. INTRODUCTION applications). Another type of BCIs relies on the detection of mental tasks (imagination of right/left hand movements, Brain Computer Interfaces (BCIs) are aimed at creating a subtraction, word association, etc…) which are detected direct communication pathway between the brain and an through slow cortical potentials (SCP) [7], readiness external device, bypassing the need for an embodiment. In potential [8] and event-related desynchronization (ERD) the last few years, research in the field of BCI has witnessed [9]. a spectacular development (see [1], [2]) and is nowadays The BCI presented here belongs to another category; it regarded as one of the most successful applications of the is based on the detection of the P300 Event-Related neurosciences. Indeed, such systems can provide a Potential (ERP: stereotyped electrophysiological response to significant improvement of the quality of life of an internal or external stimulus, [10]). This brain potential is neurologically impaired patients suffering of pathologies elicited in the context of an oddball paradigm: when a such as amyotrophic lateral sclerosis, brain stroke, subject performs the classification of two types of events, brain/spinal cord injury, etc… one of which is only rarely presented, the rare event will AC is supported by a specialization grant from the Agentschap elicit in the EEG an ERP with an enhanced positive-going voor Innovatie door Wetenschap en Technologie (IWT, Flemish component at a latency of about 300 ms (the P300 ERP, Agency for Innovation through Science and Technology). NC is [11]). The first spelling system based on the detection of the supported by IST-2007-217077. NVM is supported by IST-2004- P300 was introduced in 1988 by Farwell and Donchin [12]. 027017. JAKS acknowledges support of FWO G.0588.09, This application is nowadays one of the most studied BCI G.0302.07, CoE EF/05/006, GOA-MANET, IUAP DYSCO. MVH and the work presented here treats of this specific system. is supported by EF 2005, CREA/07/027, G.0234.04, G.0588.09, The Mind Speller® allows subjects to spell words by IUAP P5/04, GOA 10/019, IST-2004-027017 and IST-2007- focusing on the desired characters shown in a matrix display 217077). The authors wish to thank Refet Firat Yazicioglu, Tom while the rows and columns of the matrix are consecutively Torfs and Chris Van Hoof from the Interuniversity Microelectronics Centre (IMEC) in Leuven for providing us with and randomly intensified (Fig.1-left). The intensification of a the wireless EEG system. row or column containing the target symbol will elicit a P300 ERP and, by detecting this ERP, the BCI is able to consists of two parts: an amplifier coupled with a wireless identify the target row and column and thus retrieve the transmitter and a USB stick receiver (Fig. 2a, c). The data symbol the subject has in mind. are transmitted with a sampling frequency of 1 kHz for each Ideally, performing one sequence of intensifications of channel. The prototype was developed by the Interuniversity each row and column would be enough to identify the target Microelectronics Center (IMEC, [24]). We used a brain-cap symbol. Unfortunately, the low signal-to-noise ratio of the with large filling holes and sockets for active Ag/AgCl P300 ERP makes this potential almost undetectable in single electrodes (ActiCap, Brain Products, Fig. 2d). trial. The common practice is to repeat several times the The recordings were collected from eight electrodes in the sequence of intensifications, in order to average the EEG frontal, central and parietal areas, namely in positions Fz, responses and increase the signal-to-noise ratio. Depending FCz, Cz, CPz, CP1, CP2, P3, Pz, P4, according to the on the number of repetitions, this practice can lead to a international 10–20 system (Fig. 2b). The reference and dramatic increase of the time taken to communicate each ground electrodes were linked to the left and right mastoids, symbol. It is thus important to work on robust and efficient respectively (TP9, TP10). feature extraction and classification techniques to reduce this The visual stimulation consisted of a matrix of 6-by-6 number of repetitions. symbols (Fig. 1). For both training and testing stages, each An elegant way to improve the performance of a BCI is sequence of intensification consisted in the highlighting of the detection of so-called Error-related Potentials (ErrP). each row and column of the matrix only once and in random ErrPs were suggested to be generated in the anterior order. Each highlighting lasted 100 ms, followed by 100 ms cingulated cortex with a spatial distribution over fronto- of no intensification. All recordings and stimulation were central regions of the scalp and related to the subject’s performed with Matlab R2009b, the display of the stimuli perception of an error ([13], [14]). If the first studies on the and their precise timing was achieved using the presence of an ErrP in the EEG were dealing with brain Psychophysics Toolbox Extensions ([25], [26]). responses to errors made by the subject himself ([15], [16]), more recent work discusses the presence of such potential in the context of a BCI, when the user realizes that the interface failed to recognize properly his intention ([14], [17]-[20]). This latter definition is what we will refer to as ErrP in the article. In [18], the observation of an ErrP is obtained in the context of a vertical cursor controlled with mu/beta waves, while in [17] it is done in the context of a simulated BCI, (a) (b) (c) (d) where the subject manually delivers command to move a Fig. 2. (a) Wireless 8 channels amplifier. (b) Locations of the horizontal cursor. This latter experiment was then electrodes on the scalp. (c) USB stick receiver. (d) Active successfully reproduced where the BCI was still simulated electrode. with an a priori error rate but this time the subjects were using movement imagination ([14]). To our knowledge only 2.2. Experiment Design researchers from the Politecnico di Milano University ([19], [20]) recently presented some work on the error potential in Six healthy subjects (4 male, 2 female, age 22-34, 5 right the context of a P300 Speller. handed and 1 left handed) participated in the experiment. This paper reports on a study performed in our Each experiment lasted between 1h30 and 2h30, everything laboratory where 6 subjects were tested on the P300 Mind was done to keep the subject fully concentrated and the Speller® developed by our group ([21]-[23]) and for which experiments were stopped when the participants were the EEG responses to correct and incorrect feedback starting to feel tired. (moment where the BCI displays what was identified as the The first step of the experiment was to familiarize the target symbol, see Fig.1-right) were recorded during several subjects with the Mind Speller® BCI and to train the system typing sessions. We compare our results to the studies to recognize their P300 ERP. So, preliminary to any "mind- previously mentioned and discuss the possibilities of typing" performance, we performed a training session during detection of the ErrP and the ways to include it in our Mind which the participants were asked to focus consecutively on Speller® BCI. 8 symbols randomly selected by the interface. An indication of the symbol to focus on was first presented to the subject, 2. ACQUISITION OF THE DATA then the random sequence of intensifications of all the rows and columns was repeated 10 times and finally the symbol 2.1. Material was presented to the subject in the middle of the screen for 4 seconds (feedback moment, Fig. 1-right). This was repeated The EEG recordings were performed using a prototype of an for all 8 symbols. ultra low-power 8-channels wireless EEG system, which Based on the data recorded during the training session, we build the classifier for the detection of the P300 ERP. The signals were beforehand filtered between 0.3 and 15 Hz (3rd order Butterworth filter), their mean was subtracted and they were cut into 800 ms epochs starting from each stimulus onset. Those epochs were then "average- downsampled" to 80 data points (each new data point corresponds the average of the signal over its 10ms surrounding window) and finally, the data of the same classes were averaged over the desired number of trials (corresponding to the desired number of repetitions of the sequence of intensification for the spelling mode). For each trial (stimulus), we thus have 8 channels × 80 data points = 640 features to classify as a response to either a target stimulus or a non-target stimulus. A linear Support Vector Machine associated with a 10-fold cross-validation and a linesearch for the optimization of the regularization parameter was built on those training features. Training the linear SVM on 2000 data points with the modified finite Newton method proposed in [27] typically took around one minute. The second step of the experiment was to have the subjects to use Mind Speller BCI with the previously built classifier applied online to detect the P300 ERP and identify the target symbol. They would first use the system with 10 repetitions of the sequence of intensifications, in order to make them confident about the accuracy of the system. Most Table 1. Details of the performances of each participant. The 3 last of them typed their first word with almost no mistake (see columns detail for each session, the number of repetitions of the last column in Table 1). As the aim was to record EEG sequence of intensifications used to communicate each symbol, the responses to erroneous feedback, we then reduced this size of the word typed by the subject and the number of wrongly number of repetition to 5, 4 and even 3 according to how typed symbols. The fourth and fifth column summarize all the typing sessions of each subject. well the subjects were typing. Each subject spelled between 32 and 65 symbols with a number of errors comprised ms after the feedback. All three studies show quite different between 6 and 19 (see Table 1). results concerning the shape of the ErrP. It seems thus that the shape and latency times of this potential really depends 3. PRESENCE OF AN ERROR POTENTIAL on the kind of paradigm it is associated with. In [18], the subject is trained to control a cursor with his mu/beta waves; The averaged EEG responses to correct and incorrect this might already affect the shape of the ErrP. Moreover, feedback for each subject at electrode FCz and the grand the responses are recorded when the cursor hits the target (or average over subjects for each electrode are plotted in the non-target) and not when the cursor moves in the Figures 3 and 4. We also plotted the error-minus-correct intended (not-intended) direction contrarily to [14], [17] difference potentials (difference between averaged responses where the EEG responses to each cursor movement are to erroneous feedback and averaged responses to correct recorded. Also, the latency time between feedback moments feedback). may also have an influence; if in [14], [17], the EEG In their study, Schalk et al. ([18]) observed a error- feedbacks are recorded after each movement of the cursor minus-correct difference consisting of a positive potential (every 2 seconds), in [19], [20], they are recorded after each that picked about 180 ms followed by a negative potential (4 apparition of what the BCI classified as the target letter subjects were tested). In [14] and [17], this difference was (every 15 seconds). Finally, the nature of the task itself eliciting a first positive peak at 200ms after the feedback, might also influence, if a cursor moving task could involve followed by two larger negative and positive peaks at about the motor area of the brain, this is not the case of a P300 250 ms and 320 ms and a wider negative peak at 450 ms oddball paradigm. after the feedback (5 subjects were tested). Finally in [19] In our case, when looking at the grand average error- and [20] (2 and 5 subjects tested, respectively), this error- minus-correct, we can observe a negative peak followed by minus-correct difference showed a negative peak occurring positive one at about 320 ms and 450 ms respectively. Those at about 300 ms followed by a positive peak at around 400 potentials are the most prominent at the electrode sites Fz S1 S2 S3 S1 S2 S3 100 100 100 100 100 100 0 0 0 0 0 0 Amplitude (µV) Amplitude (µV) -100 -100 -100 -100 -100 -100 500 1000 500 1000 500 1000 500 1000 500 1000 500 1000 S4 S5 S6 S4 S5 S6 100 100 100 100 100 50 0 0 0 0 0 0 -100 -100 -100 -100 -100 -50 500 1000 500 1000 500 1000 500 1000 500 1000 0 500 1000 Time (ms) Time (ms) Fig. 3. EEG responses for each subject at electrode location FCz for 1 second from the feedback onset. Left: EEG responses averaged over all the correct (green) and erroneous (red) feedbacks. Right: averaged error-minus-correct. Fz FCz Cz CPz Fz FCz Cz CPz 200 200 200 200 200 200 200 200 0 0 0 0 0 0 0 0 Amplitude (µV) Amplitude (µV) -200 -200 -200 -200 -200 -200 -200 -200 0 500 1000 0 500 1000 0 500 1000 0 500 1000 0 500 1000 0 500 1000 0 500 1000 0 500 1000 CP1 CP2 P3 Pz CP1 CP2 P3 Pz 200 200 200 200 100 100 100 100 0 0 0 0 0 0 0 0 -200 -200 -200 -200 -100 -100 -100 -100 0 500 1000 0 500 1000 0 500 1000 0 500 1000 0 500 1000 0 500 1000 0 500 1000 0 500 1000 Time (ms) Time (ms) Fig. 4. EEG responses averaged over all subjects at each electrode location for 1 second from the feedback onset. Left: EEG responses averaged over all the correct (green) and erroneous (red) feedbacks. Right: averaged error-minus-correct. and FCz. Those results are in concordance with [19], [20] respect to the classification as ErrP and non-ErrP; we where a similar P300 based speller BCI as the one presented performed a permutation test at each time point (significance here was used. level 0.05, [29]). For most subject (except subject 1), the In order to assess the significance of the difference time zones corresponding to at least one of the 2 peaks between responses to erroneous feedback and responses to associated with the ErrP were statistically significant (red correct feedback we analyzed the data of each subject at the diamonds on Fig. 5-left). The same study was performed, electrode location FCz. We first "average-downsampled" regrouping this time the data from all subjects together and the signals from 1000Hz to 100Hz. Then, for each time step the coefficient of determination for both time zones i, and for all trials of a given subject, we calculated the appeared statistically significant (Fig. 5-right). coefficient of determination (square of the correlation If those results suggest an apparent discriminability coefficient, [28]) indicating the fraction of the total variance between EEG responses to both kind of feedback, the high of the EEG feedback responses xki, that was explained by the variability between responses of the same type among trials class yk of the corresponding trial k (correct feedback versus and subjects indicates the necessity of training the BCI to erroneous feedback). recognize the ErrP for each subject. This, in the case of the Mind Speller®, can be problematic due the long time cov( X i , Y ) 2 required for the acquisition of a sufficient amount of training =R 2 (i ) , i ∈ 1, ntimesteps  var( X i ) var(Y ) and testing data to build and attest of the accuracy of  {xki , k ∈ 1, ntrials  , i ∈ 1, ntimesteps } classifier. And as shown by the comparison of the results Xi = with  from [14], [17]-[20], the shape of the ErrP seems closely Y= { yk ∈ {−1,1}, k ∈ 1, ntrials }  related to the type of paradigm used for the BCI; so that we should collect the training ErrP data in the exact context in Those values are plotted in Figure 5 (left). While due to which we want to detect them. the low number of trials and the low signal-to-noise ratio of the EEG signals, the values of this coefficient of 4. DISCUSSION determination remain quite low, we can still observe some peaks along the time. Some of those peaks have the same The first difficulty is thus to gather a sufficient amount of latency time as the negative and positive peaks that we training data to build a classifier that would detect the ErrP. accounted for as ErrP in the EEG feedback responses. To We intend to direct a study with several subjects and have an idea about how significant were those peaks with S1 S2 S3 0.1 0.2 0.4 0.2 Coef. of determination R² Coef. of determination R² 0.08 0.1 0.2 0.1 0 0 0 0.06 0 500 1000 0 500 1000 0 500 1000 S4 S5 S6 0.04 0.2 0.2 0.2 0.1 0.1 0.1 0.02 0 0 0 0 0 500 1000 0 500 1000 0 500 1000 0 100 200 300 400 500 600 700 800 900 1000 Time (ms) Time (ms) Fig. 5. Coefficient of determination versus time after feedback onset for each subject independently (left) and for all subjects together (right). The red diamonds indicates the values for which the permutation test gives a p-value lower than 0.05. repeated sessions to assess what would be a sufficient If, with this strategy, not all errors are corrected, a quite amount of training data. From a practical point of view important number of them are and it presents the serious performing hours of training session in order to build an advantage of minimizing the time taken for the correction. ErrP classifier is not acceptable for a commercial device. A Looking at Table 2, one can see that more than half (17/33) solution (also proposed in [19]) would be to use the of the remaining uncorrected symbols are ranked in the third backspace symbol as a label for the presence of an ErrP in position by our classifier. One might wonder whether it the feedback response to the previous symbol and activate would be worth iterating this process when the symbol the ErrP detection once a sufficient amount of training data ranked as second appears to be wrong with the third symbol, is reached. This would still allow the user to utilize the BCI and so on, until the correct symbol is selected. That would and be familiar with the device before enhancing it with the necessitate the correct detection of several consecutive ErrP detection tool. ErrPs. Not only the theoretical probability of detection of the The second problem is the building of the classifier; correct symbol will decrease proportionally to the rank of indeed contrarily to the P300 ERP, where the signals are this symbol, but it is also possible that such ErrP would not averaged, the ErrP needs to be detected in single trial. The be elicited several times in a row. One should check the detection algorithm would have then to be robust enough to shape of the EEG response in such a case before considering overcome the low signal-to-noise ratio inherent to EEG such iterative process. Nevertheless our data suggest that it recordings and able to incrementally learn to detect an ErrP. would not be worth checking after the third ranked symbol. As we want to avoid any false detection of this error Two more possibilities unrelated to any ErrP for potential, the classifier would also have to be strongly biased improving this system can be thought of. First, one can towards the minimization of the number of false positives. consider weighting the scores of each symbol with an a Despite all these constraints, the Mind Speller® can still priori probability of occurrence given the previous symbol greatly benefit from the detection of ErrPs. Assuming that and the typing language (e.g. Dasher, [30]). Another such a tool is developed, the question arises as to how this possibility would be the use of a dictionary to automatically could be incorporated in our BCI. To detect the target letter, correct the word once it is typed. Both ideas can perfectly be the classification algorithm computes a score for each row combined and offer the advantage of not increasing the time and column of the matrix and then selects the best row and taken to communicate a word, but they would be language column. From those scores we can deduce a ranking of all specific and not usable for proper nouns or non-text based the symbols of the matrix. One strategy could be to simply communication (e.g. icon-based communication). repeat the sequence of intensifications, with eventually a lower number of repetitions, to update this ranking. But this 5. CONCLUSION would lead to an important increase in the time taken to communicate the symbol. Another strategy would then be, A first step towards the integration of ErrP detection in when the presence of an ErrP is detected, to select the the P300 Mind Speller® BCI was presented. Besides the second best letter according to the classifier’s ranking. This undeniable practical advantages of ErrP detection, the is supported by the fact that is many cases of wrong symbol necessity of single trial detection, the strong noisy detection we could observe in our experiments that at least component of EEG signals, the high inter-subject variability the column or the row of the target symbol was correctly and the paradigm dependency of this ERP make this task identified. In Table 2 is presented how such a strategy could very challenging. An appropriate way to combine it with the improve the performance of the Mind Speller® for our 6 Mind Speller® BCI has to be studied, and one should not subjects. We can observe a substantial reduction of the forget about other ways of improving the system by already number of errors (more than 3 times less for some subjects). known techniques. Table 2. Details of the possible improvement for each subject the case of a 100% accurate ErrP detection. The 7th column represents the amount of remaining wrong characters assuming that when an ErrP occurs, the symbol ranked in the second position by our classifier is selected as target symbol. The last column details the ranking of the remaining uncorrected symbols Rates," Artificial Neural Networks — ICANN 2002, Berlin, Heidelberg: Springer Berlin Heidelberg, 2002, pp. 1137-1143. REFERENCES [17] P.W. Ferrez and J. del R Millan, "Error-related EEG potentials generated during simulated brain-computer interaction," IEEE [1] "Is this the bionic man?" Nature, vol. 442, 2006, p. 109. transactions on bio-medical engineering, vol. 55, 2008, pp. 923-9. [2] P. Sajda, K. Muller, and K. Shenoy, "Brain-Computer Interfaces [18] G. Schalk, J.R. Wolpaw, D.J. McFarland, and G. Pfurtscheller, "EEG- [from the guest editors]," IEEE Signal Processing Magazine, vol. 25, based communication: presence of an error potential," Clinical 2008, pp. 16-17. neurophysiology : official journal of the International Federation of [3] B. Pesaran, S. Musallam, and R.A. Andersen, "Cognitive neural Clinical Neurophysiology, vol. 111, 2000, pp. 2138-44. prosthetics," Current biology: CB, vol. 16, 2006, pp. R77-80. [19] B. Dal Seno, M. Matteucci, and L. Mainardi, "Online Detection of [4] J.J. Vidal, "Toward direct brain-computer communication," Annual P300 and Error Potentials in a BCI Speller," Computational review of biophysics and bioengineering, vol. 2, 1973, pp. 157-80. intelligence and neuroscience, 2010, p. 307254. [5] P. Martinez, H. Bakardjian, and A. Cichocki, "Fully Online [20] G. Visconti, B. Dal Seno, M. Matteucci, and L. Mainardi, Multicommand Brain-Computer Interface with Visual Neurofeedback "Automatic Recognition of Error Potentials in a P300-Based Brain- Using SSVEP Paradigm," Computational intelligence and Computer Interface," in Proceedings of the 4th International Brain- neuroscience, vol. 2007, 2007, p. 94561. Computer Interface Workshop & Training Course, Graz: 2008, p. [6] M. Cheng, X. Gao, S. Gao, and D. Xu, "Design and implementation 238–243. of a brain-computer interface with high transfer rates," IEEE [21] N. Chumerin, N.V. Manyakov, A. Combaz, J.A. Suykens, and M.M. transactions on bio-medical engineering, vol. 49, 2002, pp. 1181-6. Van Hulle, "An application of feature selection to on-line P300 [7] N. Birbaumer, A. Kübler, N. Ghanayim, T. Hinterberger, J. detection in brain-computer interface," 2009 IEEE International Perelmouter, J. Kaiser, I. Iversen, B. Kotchoubey, N. Neumann, and Workshop on Machine Learning for Signal Processing, Ieee, 2009, H. Flor, "The thought translation device (TTD) for completely pp. 1-6. paralyzed patients," IEEE transactions on rehabilitation engineering [22] N. Chumerin, N.V. Manyakov, A. Combaz, J.A. Suykens, R.F. : a publication of the IEEE Engineering in Medicine and Biology Yazicioglu, T. Torfs, P. Merken, H.P. Neves, C.V. Hoof, and M.M. Society, vol. 8, 2000, pp. 190-3. Van Hulle, "P300 detection based on Feature Extraction in on-line [8] B. Blankertz, G. Dornhege, M. Krauledat, K. Müller, and G. Curio, Brain-Computer Interface," KI 2009: Advances in Artificial "The non-invasive Berlin Brain-Computer Interface: fast acquisition Intelligence, 2009, pp. 339-346. of effective performance in untrained subjects," NeuroImage, vol. 37, [23] A. Combaz, N.V. Manyakov, N. Chumerin, J.A. Suykens, and M.M. 2007, pp. 539-50. Van Hulle, "Feature Extraction and Classification of EEG Signals for [9] J.R. Wolpaw, D.J. McFarland, and T.M. Vaughan, "Brain-computer Rapid P300 Mind Spelling," 2009 International Conference on interface research at the Wadsworth Center," IEEE Transactions on Machine Learning and Applications, IEEE, 2009, pp. 386-391. Rehabilitation Engineering, vol. 8, 2000, pp. 222-226. [24] R. Yazicioglu, P. Merken, R. Puers, and C. Van Hoof, "Low-Power [10] S.J. Luck, An Introduction to the Event-Related Potential Technique, Low-Noise 8-Channel EEG Front-End ASIC for Ambulatory MIT Press, 2005. Acquisition Systems," 2006 Proceedings of the 32nd European [11] W.S. Pritchard, "Psychophysiology of P300," Psychological bulletin, Solid-State Circuits Conference, IEEE, 2006, pp. 247-250. vol. 89, 1981, pp. 506-40. [25] D.H. Brainard, “The Psychophysics Toolbox,” Spatial Vision, vol. [12] L.A. Farwell and E. Donchin, "Talking off the top of your head: 10, 1997, pp. 433-436. toward a mental prosthesis utilizing event-related brain potentials," [26] D.G. Pelli, “The VideoToolbox software for visual psychophysics: Electroencephalography and clinical neurophysiology, vol. 70, Transforming numbers into movies,” Spatial Vision, vol. 10, 1997, 1988, pp. 510-23. pp. 437-442. [13] C.B. Holroyd and M.G. Coles, "The neural basis of human error [27] S. Keerthi and D. DeCoste, "A modified finite Newton method for processing: reinforcement learning, dopamine, and the error-related fast solution of large scale linear SVMs," Journal of Machine negativity," Psychological review, vol. 109, 2002, pp. 679-709. Learning Research, vol. 6, 2006, pp. 341 - 361. [14] P.W. Ferrez and J. del R. Millán, "EEG-Based Brain-Computer [28] I. Guyon and A. Elisseeff, "An introduction to variable and feature Interaction: Improved Accuracy by Automatic Single-Trial Error selection," The Journal of Machine Learning Research, vol. 3, 2003, Detection," Advances in Neural Information Processing Systems, vol. p. 1157–1182. 20, 2007, pp. 1-8. [29] D.S. Moore and G. P. McCabe, Introduction to the Practice of [15] W.J. Gehring, M.G. Coles, D.E. Meyer, and E. Donchin, "A brain Statistics, 5th ed., W. H. Freeman, New York, 2006. potential manifestation of error-related processing," [30] D.J. Ward and D.J.C. MacKay, “Fast Hands-free Writing by Gaze Electroencephalography and clinical neurophysiology. Supplement, Direction,” Nature, vol. 418, August 2002, pp. 838. vol. 44, 1995, pp. 261-72. [16] G. Dornhege, G. Curio, and B. Blankertz, "Single Trial Detection of EEG Error Potentials: A Tool for Increasing BCI Transmission