{"title": "An Auditory Paradigm for Brain-Computer Interfaces", "book": "Advances in Neural Information Processing Systems", "page_first": 569, "page_last": 576, "abstract": null, "full_text": "                An Auditory Paradigm for\n               BrainComputer Interfaces\n\n\n           N. Jeremy Hill1 T. Navin Lal1 Karin Bierig1\n                                 ,                     ,\n\n             Niels Birbaumer2 and Bernhard Sch\n                                                                   olkopf1\n\n                   1 Max Planck Institute for Biological Cybernetics,\n                    Spemannstrae 38, 72076 T\n                                                ubingen, Germany.\n                     {jez|navin|bierig|bs}@tuebingen.mpg.de\n\n                        2 Institute for Medical Psychology and\n                  Behavioural Neurobiology, University of T\n                                                             ubingen,\n                     Gartenstrae 29, 72074 T\n                                               ubingen, Germany.\n                         niels.birbaumer@uni-tuebingen.de\n\n\n\n                                      Abstract\n\n        Motivated by the particular problems involved in communicating\n        with \"locked-in\" paralysed patients, we aim to develop a brain-\n        computer interface that uses auditory stimuli.            We describe a\n        paradigm that allows a user to make a binary decision by focusing\n        attention on one of two concurrent auditory stimulus sequences.\n        Using Support Vector Machine classification and Recursive Chan-\n        nel Elimination on the independent components of averaged event-\n        related potentials, we show that an untrained user's EEG data can\n        be classified with an encouragingly high level of accuracy. This\n        suggests that it is possible for users to modulate EEG signals in a\n        single trial by the conscious direction of attention, well enough to\n        be useful in BCI.\n\n\n1    Introduction\n\nThe aim of research into brain-computer interfaces (BCIs) is to allow a person to\ncontrol a computer using signals from the brain, without the need for any muscular\nmovement--for example, to allow a completely paralysed patient to communicate.\nTotal or near-total paralysis can result in cases of brain-stem stroke, cerebral palsy,\nand Amytrophic Lateral Sclerosis (ALS, also known as Lou Gehrig's disease). It\nhas been shown that some patients in a \"locked-in\" state, in which most cognitive\nfunctions are intact despite complete paralysis, can learn to communicate via an\ninterface that interprets electrical signals from the brain, measured externally by\nelectro-encephalogram (EEG) [1]. Successful approaches to such BCIs include using\nfeedback to train the patient to modulate slow cortical potentials (SCPs) to meet a\nfixed criterion [1], machine classification of signals correlated with imagined muscle\nmovements, recorded from motor and pre-motor cortical areas [2, 3], and detection\nof an event-related potential (ERP) in response to a visual stimulus event [4].\n\nThe experience of clinical groups applying BCI is that different paradigms work to\n\n\f\nvarying degrees with different patients. For some patients, long immobility and the\ndegeneration of the pyramidal cells of the motor cortex may make it difficult to\nproduce imagined-movement signals. Another concern is that in very severe cases,\nthe entire visual modality becomes unreliable: the eyes cannot adjust focus, the\nfovea cannot be moved to inspect different locations in the visual scene, meaning\nthat most of a given image will stimulate peripheral regions of retina which have\nlow spatial resolution, and since the responses of retinal ganglion cells that form the\ninput to the visual system are temporally band-pass, complete immobility of the eye\nmeans that steady visual signals will quickly fade [5]. Thus, there is considerable\nmotivation to add to the palette of available BCI paradigms by exploring EEG\nsignals that occur in response to auditory stimuli--a patient's sense of hearing is\noften uncompromised by their condition.\n\nHere, we report the results of an experiment on healthy subjects, designed to develop\na BCI paradigm in which a user can make a binary choice. We attempt to classify\nEEG signals that occur in response to two simultaneous auditory stimulus streams.\nTo communicate a binary decision, the subject focuses attention on one of the two\nstreams, left or right. Hillyard et al. [6] and others reported in the 60's and 70's that\nselective attention in a dichotic listening task caused a measurable modulation of\nEEG signals (see [7, 8] for a review). This modulation was significant when signals\nwere averaged over a large number of instances, but our aim is to discover whether\nsingle trials are classifiable, using machine-learning algorithms, with a high enough\naccuracy to be useful in a BCI.\n\n\n2    Stimuli and methods\n\nEEG signals were recorded from 15 healthy untrained subjects (9 female, 6 male)\nbetween the ages of 20 and 38, using 39 silver chloride electrodes, referenced to\nthe ears. An additional EOG electrode was positioned lateral to and slightly below\nthe left eye, to record eye movement artefacts--blinks and horizontal and vertical\nsaccades all produced clearly identifiable signals on the EOG channel. The signals\nwere filtered by an analog band-pass filter between 0.1 and 40 Hz, before being\nsampled at 256 Hz.\n\nSubjects sat 1.5m from a computer monitor screen, and performed eight 10-minute\nblocks each consisting of 50 trials. On each trial, the appearance of a fixation point\non screen was followed after 1 sec by an arrow pointing left or right (25 left, 25\nright in each block, in random order). The arrow disappeared after 500 msec, after\nwhich there was a pause of 500 msec, and then the auditory stimulus was presented,\nlasting 4 seconds. 500 msec after the end of the auditory stimulus, the fixation point\ndisappeared and there was a pause of between 2 and 4 seconds for the subject to\nrelax. While the fixation point was present, subjects were asked to keep their gaze\nfixed on it, to blink as little as possible, and not to swallow or make any other\nmovements (we wished to ensure that, as far as possible, our signals were free of\nartefacts from signals that a paralysed patient would be unable to produce).\n\nThe auditory stimulus consisted of two periodic sequences of 50-msec-long square-\nwave beeps, one presented from a speaker to the left of the subject, and the other\nfrom a speaker to the right. Each sequence contained \"target\" and \"non-target\"\nbeeps: the first three in the sequence were always non-targets, after which they could\nbe targets with independent probability 0.3. The right-hand sequence consisted of\neight beeps of frequencies 1500 Hz (non-target) and 1650 Hz (target), repeating\nwith a period of 490 msec. The left-hand sequence consisted of seven beeps of\nfrequencies 800 Hz (non-target) and 880 Hz (target), starting 70 msec after start of\nthe right-hand sequence and repeating with a period of 555 msec.\n\n\f\nAccording to the direction of the arrow on each trial, subjects were instructed to\ncount the number of target beeps in either the left or right sequence. In the pause\nbetween trials, they were instructed to report the number of target beeps using a\nnumeric keypad.1\n\n                                                                                                                            deviant tones\n                                                                                                                (longer duration, or absent)\n                                                                                             time (sec)\n                                                             0         0.5    1       1.5      2    2.5    3         3.5    4\n                                                     +1\n\n\n\n\n\n                                                     +0.5\n     concatenated       ed                                                                                                       LEFT and RIGHT\n       averaged               u                                                                                                   sound signals of\n                                   lit                  0\n        signal                            p                                                                                       different periods\n                                               am\n                                                     -0.5\n\n\n\n\n\n                                                       -1\n\n\n\n\n\n                                                                  1\n                                                                                                                1\n                                                                  n                (epoch averaging)            n\n\n\n\n\n\nFigure 1: Schematic illustration of the acoustic stimuli used in the experiment, and\nof the averaging process used in preprocessing method A (illustrated by showing\nwhat would happen if the sound signals themselves were averaged)\n\nThe sequences differed in location and pitch in order to help the subjects focus\ntheir attention on one sequence and ignore the other. The task of reporting the\nnumber of target beeps was instituted in order to keep the subjects alert, and to\nmake the task more concrete, because subjects in pilot experiments found that just\nbeing asked \"listen to the left\" or \"listen to the right\" was too vague a task demand\nto perform well over the course of 400 repetitions.2 The regular repetition of the\nbeeps, at the two different periods, was designed to allow the average ERP to a left-\nhand beep on a single trial to be examined with minimal contamination by ERPs\nto right-hand beeps, and vice versa: figure 2 illustrates that, when the periods of\none signal are averaged, signals correlated with that sequence add in phase, whereas\nsignals correlated with the other sequence spread out, out of phase. Comparison\nof the average response to a left beat with the average response to a right beat,\non a single trial, should thus emphasize any modulating effect of the direction of\nattention on the ERP, of the kind described by Hillyard et al. [6].\n\n\n   1 In order to avoid contamination of the EEG signals with movement artefacts, a few\npractice trials were performed before the first block, so that subjects learned to wait until\nthe fixation point was out before looking at the keypad or beginning the hand movement\ntoward it.\n   2 Although a paralysed patient would clearly be unable to give responses in this way, it\nis hoped that this extra motivation would not be necessary.\n\n\f\nAn additional stimulus feature was designed to investigate whether mismatch neg-\nativity (MMN) could form a useful basis for a BCI. Mismatch negativity is a differ-\nence between the ERP to standard stimuli and the ERP to deviant stimuli, i.e. rare\nstimulus events (with probability of occurrence typically around 0.1) which differ\nin some manner from the more regular standards. MMN is treated in detail by\nNa\n     atanen [9]. It has been associated with the distracting effect of the occurrence of\na deviant while processing standards, and while it occurs to stimuli outside as well\nas inside the focus of attention, there is evidence to suggest that this distraction\neffect is larger the more similar the (task-irrelevant) deviant stimulus is to the (task-\nrelevant) standards [10]. Thus there is the possibility that a deviant stimulus (say,\na longer beep) inserted into the sequence to which the subject is attending (same\nside, same pitch) might elicit a larger MMN signal than a deviant in the unattended\nsequence. To explore this, after at least two standard beats of each trial, one of the\nbeats (randomly chosen, with the constraint that the epoch following the deviant\non the left should not overlap with the epoch following the deviant on the right)\nwas made to deviate on each trial. (Note the frequencies of occurrence of the de-\nviants were 1/7 and 1/8 rather than the ideal 1/10: the double contraint of having\nmanageably short trials and a reasonable epoch length meant that the number of\nbeeps in the left and right sequences was limited to seven and eight respectively,\nand clearly to use MMN in BCI, every trial has to have at least one deviant in each\nsequence.) For 8 subjects, the deviant beat was simply a silent beat--a disruptive\npause in the otherwise regular sequence. For the remaining 7 subjects, the deviant\nbeat was a beep lasting 100 msec instead of the usual 50 msec (as in the distraction\nparadigm of Schr\n                    oger and Wolff [10], the difference between deviant and standard is\non a task-irrelevant dimension--in our case duration, the task being to discriminate\npitch). A sixteenth subject, in the long-deviant condition, had to be eliminated\nbecause of poor signal quality.\n\n\n3       Analysis\n\nAs a first step in analyzing the data, the raw EEG signals were examined by eye for\neach of the 400 trials of each of the subjects. Trials were rejected if they contained\nobvious large artefact signals caused by blinks or saccades (visible in the EOG and\nacross most of the frontal positions), small periodic eye movements, or other muscle\nmovements (neck and brow, judged from electrode positions O9 and O10, Fp1, Fpz\nand Fp2). Between 6 and 228 trials had to be rejected out of 400, depending on\nthe subject.\n\nOne of two alternative preprocessing methods was then used. In order to look for\neffects of the attention-modulation reported by Hillyard et al, method (A) took the\naverage ERP in response to standard beats (discarding the first beat). In order to\nlook for possible attention-modulation of MMN, method (B) subtracted the average\nresponse to standards from the response to the deviant beat. In both methods, the\naverage ERP signal to beats on the left was concatenated with the average ERP\nsignal following beats on the right, as depicted in figure 2 (for illustrative purposes\nthe figure uses the sound signal itself, rather than an ERP). For each trial, either\npreprocessing method resulted in a signal of 142 (left) + 125 (right) = 267 time\nsamples for each of 40 channels (39 EEG channels plus one EOG), for a total of\n10680 input dimensions to the classifier.\n\nThe classifier used was a linear hard-margin Support Vector Machine (SVM) [11].\nTo evaluate its performance, the trials from a single subject were split into ten non-\noverlapping partitions of equal size: each such partition was used in turn as a test\nset for evaluating the performance of the classifier trained on the other 90% of the\n\n\f\ntrials. Before training, linear Independent Component Analysis (ICA) was carried\nout on the training set in order to perform blind source separation--this is a common\ntechnique in the analysis of EEG data [12, 13], since signals measured through the\nintervening skull, meninges and cerebro-spinal fluid are of low spatial resolution,\nand the activity measured from neighbouring EEG electrodes can be assumed to be\nhighly correlated mixtures of the underlying sources. For the purposes of the ICA,\nthe concatenation of all the preprocessed signals from one EEG channel, from all\ntrials in the training partition, was treated as a single mixture signal. A 40-by-40\nseparating matrix was obtained using the stabilized deflation algorithm from version\n2.1 of FastICA [14]. This matrix, computed only from the training set, was then\nused to separate the signals in both the training set and the test set. Then, the\nsignals were centered and normalized: for each averaged (unmixed) ERP in each of\nthe 40 ICs of each trial, the mean was subtracted, and the signal was divided by its\n2-norm. Thus the entry Kij in the kernel matrix of the SVM was proportional to the\nsum of the coefficients of correlation between corresponding epochs in trials i and j.\nThe SVM was then trained and tested. Single-trial error rate was estimated as the\nmean proportion of misclassified test trials across the ten folds. For comparison,\nthe classification was also performed on the mixture signals without ICA, and with\nand without the normalizing step.\n\nResults are shown in table 1. Due to space constraints, standard error values\nfor the estimated error rates are not shown: standard error was typically 0.025,\nand maximally 0.04. It can be seen that the best error rate obtainable with\na given subject varies according to the subject, between 3% and 37%, in a way\nthat is not entirely explained by the differences in the numbers of good (artefact-\nfree) trials available. ICA generally improved the results, by anything up to 14%.\nPreprocessing method (B) generally performed poorly (minimum 19% error, and\ngenerally over 35%). Any attention-dependent modulation of an MMN signal is\napparently too small relative to the noise (signals from method B were generally\nnoisier than those from method A, because the latter, averaged over 5 or 6 epochs\nwithin a trial, are subtracted from signals that come from only one epoch per\ntrial in order to produce the method B average). For preprocessing method A,\nnormalization generally produced a small improvement.\n\nThus, promising results can be obtained using the average ERP in response to\nstandard beeps, using ICA followed by normalization (fourth results column): error\nrates of 515% for some some subjects are comparable with the performance of, for\nexample, well-trained patients in an SCP paradigm [1], and correspond to informa-\ntion transfer rates of 0.40.7 bits per trial (say, 47 bits per minute). Note that,\ndespite the fact that this method does not use the ERPs that occur in response\nto deviant beats, the results for subject in the silent-deviant condition were gener-\nally better than for those in the long-deviant condition. It may be that the more\nirregular-sounding sequences with silent beats forced the subjects to concentrate\nharder in order to perform the counting task--alternatively, it may simply be that\nthis group of subjects could concentrate less well, an interpretation which is also\nsuggested by the fact that more trials had to be rejected from their data sets).\n\nIn order to examine the extent to which the dimensionality of the classification\nproblem could be reduced, recursive feature elimination [15] was performed (limited\nnow to preprocessing method A with ICA and normalization). For each of ten folds,\nICA and normalization was performed, then an SVM was trained and tested. For\neach independent component j, an elimination criterion value cj =                w2 was\n                                                                         iFj     i\ncomputed, where w is the hyperplane normal vector of the trained SVM, and Fj\nis the set of indices to features that are part of component j. The IC with the\nlowest criterion score cj was deemed to be the least influential for classification,\n\n\f\nTable 1: SVM classification error rates: the best rates for each of the preprocessing\nmethods, A and B (see text), are in bold. The symbol                                 is used to denote\nnormalization during pre-processing as described in the text, and the symbol  is\nused to denote no normalization.\n           deviant      #                Method A                               Method B\n  subj.    duration    good        no ICA              ICA           no ICA                     ICA\n           (msec)      trials                                                                    \n   CM         0        326       0.08    0.06    0.06    0.04      0.36         0.35      0.26    0.25\n   CN         0        250       0.26    0.19    0.28    0.14      0.43         0.44      0.38         0.40\n   GH         0        198       0.34    0.27    0.35    0.22      0.41         0.41      0.39         0.43\n   JH         0        348       0.21    0.19    0.14    0.08      0.31         0.42      0.28         0.35\n   KT         0        380       0.23    0.21    0.15    0.07      0.41         0.36      0.35    0.34\n  KW          0        394       0.18    0.14    0.06    0.03      0.34         0.39      0.19         0.23\n   TD         0        371       0.22    0.18    0.15    0.10      0.35         0.39      0.29    0.28\n   TT         0        367       0.32    0.31    0.33    0.32      0.40         0.42      0.39         0.43\n   AH        100       353       0.22    0.22    0.17    0.16      0.41    0.41           0.45         0.46\n   AK        100       172       0.35    0.31    0.34    0.22      0.50         0.46      0.50    0.42\n   CG        100       271       0.37    0.29    0.31    0.28      0.51         0.47      0.48    0.44\n   CH        100       375       0.31    0.28    0.26    0.22      0.49         0.46      0.46    0.44\n   DK        100       241       0.34    0.34    0.35    0.30      0.45         0.44      0.42    0.40\n   KB        100       363       0.21    0.21    0.15    0.10      0.42         0.47      0.39         0.41\n   SK        100       239       0.47    0.43    0.40    0.37      0.46         0.49      0.45         0.51\n\n\n\nand the corresponding features Fj were removed. Then the SVM was re-trained\nand re-tested, and the elimination process iterated until one channel remained.\nThe removal of batches of features in this way is similar to the Recursive Channel\nElimination approach to BCI introduced by Lal et al. [3], except that independent\ncomponents are removed instead of mixtures (a convenient acronym would therefore\nbe RICE, for Recursive Independent Component Elimination).\n\nResults for the two subject groups are plotted in the left and right panels of figure 3,\nshowing estimated error rates averaged over ten folds against the number of ICs used\nfor classification. Each subject's initials, together with the number of useable trials\nthat subject performed, are printed to the right of the corresponding curve.3 It can\nbe seen that a fairly large number of ICs (around 2025 out of the 40) contribute to\nthe classification: this may indicate that the useful information in the EEG signals\nis diffused fairly widely between the areas of the brain from which we are detecting\nsignals (indeed, this is in accordance with much auditory-ERP and -MMN research,\nin which strong signals are often measured at the vertex, quite far from the auditory\ncortex [6, 7, 8, 9]). One of the motivations for reducing the dimensionality of the\ndata is to determine whether performance can be improved as irrelevant noise is\neliminated, and as the probability of overfitting decreases. However, these factors do\nnot seem to limit performance on the current data: for most subjects, performance\ndoes not improve as features are eliminated, instead remaining roughly constant\nuntil fewer than 2025 ICs remain. A possible exception is KT, whose performance\nmay improve by 23% after elimination of 20 components, and a clearer exception\n\n   3 RICE was also carried out using the full 400 trials for each subject (results not shown).\nDespite the (sometimes drastic) reduction in the number of trials, rejection by eye of\nartefact trials did not raise the classification error rate by an appreciable amount. The\none exception was subject SK, for whom the probability of mis-classification increased by\nabout 0.1 when 161 trials containing strong movement signals were removed--clearly this\nsubject's movements were classifiably dependent on whether he was attending to the left\nor to the right.\n\n\f\n                                            deviant duration = 0                deviant duration = 100 msec\n                                 0.5\n\n\n                                0.45\n\n\n                                 0.4\n\n                                                                                                               SK (239)\n\n                                0.35\n\n     te                                                             TT (367)                                   DK (241)\n           ra 0.3\n           r                                                                                                   CG (271)\n                erron 0.25\n                                                                                                               AK (172)\n                catio                                               GH (198)                                   CH (375)\n                         ssifi 0.2\n                         cla\n                                                                                                               AH (353)\n                                0.15                                CN (250)\n\n                                                                    TD (371)\n                                 0.1                                                                           KB (363)\n                                                                    JH (348)\n                                                                    KT (380)\n                                0.05                                CM (326)\n                                                                    KW (394)\n\n                                  0       5 10 15 20 25 30 35 40                  5 10 15 20 25 30 35 40\n\n                                         number of ICs retained                   number of ICs retained\n\n                                        Figure 2: Results of recursive independent component elimination\n\n\nis CG, for whom elimination of 25 components yields an improvement of roughly\n10%.\n\nThe ranking returned by the RICE method is somewhat difficult to interpret, not\nleast because each fold of the procedure can compute a different ICA decomposition,\nwhose independent components are not necessarily readily identifiable with one\nanother. A thorough analysis is not possible here--however, with the mixture\nweightings for many ICs spread very widely around the electrode array, we found\nno strong evidence for or against the particular involvement of muscle movement\nartefact signals in the classification.\n\n\n4                               Conclusion\n\nDespite wide variation in performance between subjects, which is to be expected\nin the analysis of EEG data, our classification results suggest that it is possible\nfor a user with no previous training to direct conscious attention, and thereby\nmodulate the event-related potentials that occur in response to auditory stimuli\nreliably enough, on a single trial, to provide a useful basis for a BCI. The information\nused by the classifier seems to be diffused fairly widely over the scalp. While\n\n\f\nthe ranking from recursive independent component elimination did not reveal any\nevidence of an overwhelming contribution from artefacts related to muscle activity,\nit is not possible to rule out completely the involvement of such artefacts--possibly\nthe only way to be sure of this is to implement the interface with locked-in patients,\npreparations for which are underway.\n\nAcknowledgments\n\nMany thanks to Prof. Kuno Kirschfeld and Bernd Battes for the use of their labo-\nratory.\n\n\nReferences\n\n [1] N. Birbaumer, A. K\n                              ubler, N. Ghanayim, T. Hinterberger, J. Perelmouter, J. Kaiser,\n    I. Iversen, B. Kotchoubey, N. Neumann, and H. Flor. The Thought Translation De-\n    vice (TTD) for Completely Paralyzed Patients. IEEE Transactions on Rehabilitation\n    Engineering, 8(2):190193, June 2000.\n\n [2] G. Pfurtscheller., C. Neuper amd A. Schl\n                                                    ogl, and K. Lugger. Separability of EEG\n    signals recorded during right and left motor imagery using adaptive autoregressive\n    parameters. IEEE Transactions on Rehabilitation Engineering, 6(3):316325, 1998.\n [3] T.N. Lal, M. Schr\n                             oder, T. Hinterberger, J. Weston, M. Bogdan, N. Birbaumer,\n    and B. Sch\n                     olkopf. Support Vector Channel Selection in BCI. IEEE Transactions\n    on Biomedical Engineering. Special Issue on Brain-Computer Interfaces, 51(6):1003\n    1010, June 2004.\n\n [4] E. Donchin, K.M. Spencer, and R. Wijesinghe. The menatal prosthesis: Assessing the\n    speed of a P300-based brain-computer interface. IEEE Transactions on Rehabilitation\n    Engineering, 8:174179, 2000.\n\n [5] L.A. Riggs, F. Ratliff, J.C. Cornsweet, and T.N. Cornsweet. The disappearance of\n    steadily fixated visual test objects. Journal of the Optical Society of America, 43:495\n    501, 1953.\n\n [6] S.A. Hillyard, R.F. Hink, V.L. Schwent, and T.W. Picton. Electrical signs of selective\n    attention in the human brain. Science, 182:177180, 1973.\n\n [7] R. N\n           a\n              at\n                 anen. Processing negativity: an evoked-potential reflection of selective atten-\n    tion. Psychological Bulletin, 92(3):605640, 1982.\n [8] R. N\n           a\n              at\n                 anen. The role of attention in auditory information processing as revealed by\n    event-related potentials and other brain measures of cognitive function. Behavioral\n    and Brain Sciences, 13:201288, 1990.\n\n [9] R. N\n           a\n              at\n                 anen. Attention and Brain Function. Erlbaum, Hillsdale NJ, 1992.\n[10] E. Schr\n                 oger and C. Wolff. Behavioral and electrophysiological effects of task-irrelevant\n    sound change: a new distraction paradigm. Cognitive Brain Research, 7:7187, 1998.\n\n[11] B. Sch\n              olkopf and A. Smola. Learning with Kernels. MIT Press, Cambridge, USA,\n    2002.\n[12] K.R. M\n                 uller, J. Kohlmorgen, A. Ziehe, and B. Blankertz. Decomposition algorithms\n    for analysing brain signals. In S. Haykin, editor, Adaptive Systems for Signal Pro-\n    cessing, Communications and Control, pages 105110, 2000.\n\n[13] A. Delorme and S. Makeig. EEGLAB: an open source toolbox for analysis of single-\n    trial EEG dynamics including Independent Component Analysis. Journal of Neuro-\n    science Methods, 134:921, 2004.\n\n[14] A. Hyv\n                 arinen. Fast and robust fixed-point algorithms for Independent Component\n    Analysis. IEEE Transactions on Neural Networks, 10(3):626634, 1999.\n[15] I. Guyon, J. Weston, S. Barnhill, and V. Vapnik. Gene Selection for Cancer Clas-\n    sification using Support Vector Machines. Journal of Machine Learning Research,\n    3:14391461, 2003.\n\n\f\n", "award": [], "sourceid": 2551, "authors": [{"given_name": "N.", "family_name": "Hill", "institution": null}, {"given_name": "Thomas", "family_name": "Lal", "institution": null}, {"given_name": "Karin", "family_name": "Bierig", "institution": null}, {"given_name": "Niels", "family_name": "Birbaumer", "institution": null}, {"given_name": "Bernhard", "family_name": "Sch\u00f6lkopf", "institution": null}]}