Examining the potential of autoencoders for automatic feature extraction in wireless cognitive load inference ; Preučevanje potenciala samokodirnikov za avtomatično grajenje značilk pri brezžičnem ugotavljanju kognitivne obremenjenosti

Understanding a user's cognitive load could enable a range of applications, from better driving assistance solutions, to educational games. This thesis analyses data from a wireless cognitive load inference study and examines the potential of autoencoders for extracting features from this data. These features would then be used for a more accurate inference of the user's cognitive load. We compare the performance of 7 classifiers using 4 handcrafted features from the original study and 4 new autoencoder based features. We define three flavours of the cognitive load inference problem and 7 subsets of data and test each combination of problem flavour, subset, classifier and feature. Our findings show that the handcrafted features perform best in most cases, however in some cases the autoencoder-based features provide a slightly higher classification accuracy. Although for one problem flavour the highest accuracies reach over 85 %, performance on the other two problems flavours have much more room for improvement, showing that inference of cognitive load from wireless signals is challenging and requires further study. ; Razumevanje uporabnikove kognitivne obremenitve lahko omogoči širok nabor uporab, kot so boljši sistemi za pomoč med vožnjo ali igre za poučevanje. Ta diplomska naloga analizira podatke iz raziskave o brezžičnem ugotavljanju kognitivne obremenitve in preuči potencial samokodirnikov za grajenje značilk iz teh podatkov. Te značilke bi bile nato uporabljene za bolj natančno ugotavljanje uporabnikove kognitivne obremenjenosti. Primerjamo zmogljivost sedmih razvrščevalnikov na podlagi štirih skupin značilk iz originalne raziskave in štirih novih skupin značilk pridobljenih s pomočjo samokodirnikov. Definiramo tri vrste problema ugotavljanja kognitivne obremenitve in sedem podmnožic podatkov ter preverimo delovanje vsake kombinacije vrste problema, podmnožice, razvrščevalnika in značilke. Naše ugotovitve kažejo, da značilke iz originalne raziskave v večini primerov dosežejo najvišjo klasifikacijsko ...