Globoko učenje in vmesniki možgani-računalnik

CASERMAN, ROK (author), Čehovin Zajc, Luka (mentor)

Vmesniki možgani-računalnik omogočajo človeku nadzor nad napravo s pomočjo merjenja šibke možganske električne aktivnosti. S tehniko, imenovano elektroencefalografija, je mogoče vzpostaviti komunikacijski kanal, preko katerega se lahko uporabnik sporazumeva z napravo. Prva večja omejitev vmesnikov, ki delujejo s pomočjo elektroencefalografije, je cena uporabljene opreme, ki velikokrat presega denarne zmožnosti posameznikov. Poleg tega pa izbor značilk možganskih signalov, ki se uporabljajo v fazi klasifikacije, zahteva obsežno domensko znanje. V okviru diplomskega dela je bil razvit vmesnik možgani-računalnik zgolj z uporabo prostodostopnih oz. nizkocenovnih orodij. Hkrati pa smo za klasifikacijo signalov uporabili konvolucijsko nevronsko mrežo, za katero izbor značilk ni potreben. Implementirani vmesnik smo sistematično ovrednotili na dveh klasifikacijskih metodah in tako pokazali, da je globoko učenje v vmesnikih možgani-računalnik pod določenimi pogoji enako ali celo boljše kot obstoječe klasifikacijske tehnike. Rezultati pa so tudi pokazali nizko klasifikacijsko točnost konvolucijske nevronske mreže v primeru, ko so bili učni podatki zajeti z drugačno opremo, kot jo uporablja implementirani vmesnik, kar nakazuje na neprenosljivost uporabljenega modela med različnimi strojnimi opremami.

• Language: Slovenian
• Keywords: vmesniki možgani-računalnik, globoko učenje, elektroencefalografija
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2023
• PID: 20.500.12556/RUL-144766
• COBISS.SI-ID: 147532803
• Publication date in RUL: 10.03.2023

Secondary language

• Language: English
• Title: Deep learning and brain-computer interfaces

Abstract

Brain-computer interfaces enable humans to control machines by measuring weak electrical activity produced by the brain. An user-machine communication channel used can be established through a technique called electroencephalography. The first bigger obstacle of such interfaces is the cost of equipment associated with them, which is in most cases out of individuals reach. Furthermore the feature selection requires a large amount of domain knowledge which can be a difficult process. In this paper we've developed a brain-computer interface using components which are either free to use or low cost. For feature selection process we've leveraged convolutional neural networks which don't require this step. The implemented interface was evaluated using two classification techniques with the result showing that under some circumstances deep learning is equivalent or better to existing classification methods. Results have also shown that using convolutional neural network, which was trained on a dataset captured with different equipment to which the implemented BCI was using, resulted in poor classification performance, which indicates non-transferability of used model between different hardware.

• Keywords: brain-computer interface, deep learning, electroencephalography