Interaktiivne kohandamisvõimeline õppimine: uus lähenemine aju-arvuti liidese süsteemi õpetamiseks
Files
Date
2013
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Tartu Ülikool
Abstract
Aju-arvuti liides (AAL) on süsteem, mis võimaldab infovahetust inimese aju ja arvuti vahel. Kasutades erinevaid neuropildistuste tehnikaid aju aktiivsust salvestatakse ja saadetakse arvutisse, kus signaal töödeldakse masinõpe meetoditega. AALi põhieesmärk on anda inimesele võimalust juhtida välisseadet kasutades mõttejõudu.
Inimese mõtteseisundite eristame on raske ülesanne, mis ei ole lahendatav ainult masinõpe kasutamisega. Vastuvõetav klassifitseerimise täpsuse tase on saavutatav pärast pikajalist õpetamise protsessi, mille jooksul inimene õpib kuidas ta peab tekitama sobivad mõtteseisundid, ning arvuti loob mudeli, mis oskab neid eristada.
Käesolevas töös me esitame uut lähenemist AAL süsteemi õpetamise protsessi jaoks. See põhineb inimese ja arvuti koostoimimise ideel, mille jooksul mõlemad osapooled adapteerivad oma käitumist vastavalt sellele, millist tagasisided nad saavad suhtlemise ajal. Pakutud viisi vastandiks on võetud traditsiooniline lähenemine, kus katseisik ei saa tagasisidet õppeprotsessi edukusest selle käigus.
Teine uudsus traditsioonilise meetodiga võrreldes on juhendamata õppealgoritmi kasutamine (iseorganiseeriv kaart, SOM) meie süsteemi tuumana. Algne iseorganiseeruva kaardi algoritm on täiendatud niimoodi, et ta esindab tõenäosusliku ennustamise mudelit, mis oskab klassifitseerida aju signaali, anda tagasisidet katseisikule ning vajadusel kohandada mudelit reaalajas.
Tuginedes läbiviidud eksperimentide tulemustel e järeldame, et interaktiivne lähenemine süsteemi õpetamiseks omab hulk eelisi traditsioonilise meetodiga võrreldes.
A Brain-Computer Interface is a system which allows communication between a human and a computer. Using various neuroimaging techniques the brain activity is recorded and transmitted to the computer, where the signal is analyzed with the help of machine learning methods. The ultimate goal of BCI is to empower the human with the ability to control the external device with the power of thought. However, distinguishing mental states of a human is a challenging task and standard machine learning alone is not enough to solve the problem. Acceptable level of performance can be achieved after a long training process, during which the human learns how to produce suitable mental states and the machine creates a model, which is able to classify the signal. In this thesis we proposed a conceptually new approach to the process of training a BCI system. It relies on the idea of the interaction between the test subject and the machine and the ability of those two agents to adapt their behavior accordingly to the information they receive during the learning process. The approach is proposed as a counterpart to the traditional BCI training, where the test subject does not receive any feedback. Another novelty in comparison to the traditional approach is using an unsupervised learning algorithm (SOM) as the core of the learning system. The original concept of self-organizing maps is amended to represent a probabilistic predictive model, which can be used to classify the brain signal, provide feedback and adapt the model in real time. Based on the results of the conducted experiments we conclude that adaptive learning process has the multiple major advantages over the traditional one.
A Brain-Computer Interface is a system which allows communication between a human and a computer. Using various neuroimaging techniques the brain activity is recorded and transmitted to the computer, where the signal is analyzed with the help of machine learning methods. The ultimate goal of BCI is to empower the human with the ability to control the external device with the power of thought. However, distinguishing mental states of a human is a challenging task and standard machine learning alone is not enough to solve the problem. Acceptable level of performance can be achieved after a long training process, during which the human learns how to produce suitable mental states and the machine creates a model, which is able to classify the signal. In this thesis we proposed a conceptually new approach to the process of training a BCI system. It relies on the idea of the interaction between the test subject and the machine and the ability of those two agents to adapt their behavior accordingly to the information they receive during the learning process. The approach is proposed as a counterpart to the traditional BCI training, where the test subject does not receive any feedback. Another novelty in comparison to the traditional approach is using an unsupervised learning algorithm (SOM) as the core of the learning system. The original concept of self-organizing maps is amended to represent a probabilistic predictive model, which can be used to classify the brain signal, provide feedback and adapt the model in real time. Based on the results of the conducted experiments we conclude that adaptive learning process has the multiple major advantages over the traditional one.