Tõhusa aju-arvuti liidese suunas: täiusliku täpsuse saavutamine aja ohverdamisega
Abstract
Aju-arvuti liides (AAL) on süsteem aju elektriliste impulside väljavõtmiseks janende kasutamiseks arvuti tarkvara juhtimiseks. AAL opereerimiseks peab kasutaja kontsentreeruma mingile mõttelisele ülesandele. Lisaks impulside mõõtmisele muudab AAL elekroonilisi signaale digitaalseks ja selle järgi tuvastab vastava arvuti käsu. Kahjuks on õige käsu tuvastamise tõenäosus alati alla 100%, mistõttu AAL süsteemide tõhusus on võrdlemisi madal.Madal tõhusus on AAL-i jaoks suureks probleemiks, sest senikaua kuni needsüsteemid pakuvad madalaid tuvastamise täpsuseid, jäävad need paljudes valdkondades ilma kasutamiseta. Antud probleemi lahendamiseks enamasti üritatakse tõsta AAL-i täpsust ühe kontsentreerumiskatse raames ja ei pöörata tähelepanu kontsentreerumiskatse kestvusele. Meie lähenemine aga põhineb arusaamisel, kui palju kontsentreerumiskatseid on vaja kasutajal järjest teostada (s.t kui kaua aega on nõutud), et saavutada 99% täpsus.Selles töös kirjeldatud lahendus põhineb Condorcet kohtu teoreemil [1]. Teoreem väidab, et kui on olemas kaks valikuvõimalust ja tõenäosus valida õiget on suurem kui 50%, kui me teostame mitu valimiskatset järjest, siis tõenäosus, et valitakse õiget valikut tõuseb iga järgneva valimiskatsega. Antud töös rakendasime põhilist Condorcet printsiipi aju-arvuti liidesele. Kõigepealt me arendame süsteemi, mis on suuteline saavutama ühe mõttelise ülesande kontsentreerumiskatse täpsuseks rohkem kui 50% ja seejärel proovime läbi mitme kontsentreerumiskatse parandada keskmist täpsust. Me eeldame, et kui kasutada piisavat kogust kontsentreerumiskatseid, siis me jõuame 99% klassifitseerimistäpsuseni. Me võrdleme teoreetilisi tulemusi eksperimentaalsetega ning arutleme nende üle. AAL tehnoloogia on võrdlemisi uus valdkond. Selle tehnoloogia täielik toomine meie igapäevaellu nõuab tugevat panust teadlastelt ja inseneridelt, et muuta AAL usaldusväärseks süsteemiks. Antud töö eesmärk on panustada AAL süsteemi kindlusesse.
Brain-computer interface (BCI) is a computer system for extracting brain electricneural signals and using them to control computer applications. For the operationBCI requires a user to concentrate on some mental tasks. Besides measuringthe signals, BCI converts raw electric signal to digital representation and maps thedata to computer commands. Unfortunately, the probability of predicting the rightcommand is below 100% and therefore the reliability of these systems is relativelylow.Low reliability is a huge problem for BCI, since they will not be widely trustedand used while the prediction accuracy is low. The existing solutions usually tryto improve the prediction accuracy of BCI without focusing too much on the timewhat is required for a single user’s concentration attempt. They apply differentprediction models and signal processing techniques in order to raise the accuracyof prediction. Our solution goes the opposite way – it tries to discover how manyconcentration attempts should be done in a row (i.e how long does it take), toguarantee the prediction accuracy of 99%.The solution described in the thesis is based on Condorcet’s jury theorem [1].It states that if we have two options and the chance to pick correct is larger than50%, then, if we make several attempts in a row, the probability to pick the correctoption by majority vote is rising with the number of attempts. In this work weapply the main Condorcet’s principle in a BCI perspective. First we develop asystem that can reach the single concentration attempt’s prediction accuracy tobe more than 50% and then we use multiple concentration attempts in a row toimprove the overall accuracy. We expect that given enough attempts we can reach99% classification accuracy. We compare the empirical results with the theoreticalestimates and discuss them.The BCI technology is a relatively young field. In order to fully integrate itinto our ordinary life, the contribution from scientists and engineers is required forconverting BCI to a reliable system. The following work contributes to reliabilityof BCI systems.
Brain-computer interface (BCI) is a computer system for extracting brain electricneural signals and using them to control computer applications. For the operationBCI requires a user to concentrate on some mental tasks. Besides measuringthe signals, BCI converts raw electric signal to digital representation and maps thedata to computer commands. Unfortunately, the probability of predicting the rightcommand is below 100% and therefore the reliability of these systems is relativelylow.Low reliability is a huge problem for BCI, since they will not be widely trustedand used while the prediction accuracy is low. The existing solutions usually tryto improve the prediction accuracy of BCI without focusing too much on the timewhat is required for a single user’s concentration attempt. They apply differentprediction models and signal processing techniques in order to raise the accuracyof prediction. Our solution goes the opposite way – it tries to discover how manyconcentration attempts should be done in a row (i.e how long does it take), toguarantee the prediction accuracy of 99%.The solution described in the thesis is based on Condorcet’s jury theorem [1].It states that if we have two options and the chance to pick correct is larger than50%, then, if we make several attempts in a row, the probability to pick the correctoption by majority vote is rising with the number of attempts. In this work weapply the main Condorcet’s principle in a BCI perspective. First we develop asystem that can reach the single concentration attempt’s prediction accuracy tobe more than 50% and then we use multiple concentration attempts in a row toimprove the overall accuracy. We expect that given enough attempts we can reach99% classification accuracy. We compare the empirical results with the theoreticalestimates and discuss them.The BCI technology is a relatively young field. In order to fully integrate itinto our ordinary life, the contribution from scientists and engineers is required forconverting BCI to a reliable system. The following work contributes to reliabilityof BCI systems.