Piltide kujutamine inimese ajutegevuses
Files
Date
2016
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Antud bakalaureusetöö uurib, kuidas on pildid kujutatud inimese ajutegevuses. Andmestik, mida töös kasutatakse, pärineb Lyoni Ülikooli haiglas tehtud uuringutest, kus mõõdeti intrakraniaalsete elektroodide abil inimese ajutegevust visuaalse ülesande ajal. Bakalaureusetöö eesmärk on analüüsida neid andmeid, täpsemalt kõrge gamma sageduse vahemikus olevat aktiivsust, juhendamata masinõppe meetoditega. Antud bakalaureusetöös kasutati erinevaid algoritme, et leida andmetest struktuuri, mis kirjeldaks elektroodide abil salvestatud visuaalse stiimuli ajal toimunud reaktsioonide sarnasusi. Lisaks analüüsiti kuivõrd on võimalik eristada pildi kategooriaid (majad, näod, loomad jne.) vastavalt tekkinud aju aktiivsusele. Kasutatud analüüsi meetoditest oli edukaim visualiseerimine. Katses esitatud pildigruppidest eristusid kõige paremini nägude kujutised. Samuti klasterdusid mingil määral majad ja skrambleeritud (segipaisatud) pildid. Elektroodide visualiseerimisel tekkisid selged klastrid, mis koosnesid aga mitmest Brodmanni piirkonnast pärit andmepunktidest. Kahendklasterdamine\n\rtähelepanuväärseid tulemusi ei andnud.
This bachelor thesis explores how images are represented in human brain activity. Data used in this thesis were collected in the University Hospital of Lyon with experiments where human neural activity was recorded with intracranial electrodes during a simple visual task. The aim of this thesis is to analyse the data, more specifically the activity in the high-gamma frequency range, with unsupervised machine learning methods to find structure in it. In particular, similarity of neural responses (recorded by electrodes) to images and the differences in activity according to image categories.\n\rVisualization and biclustering were used in this thesis. Out of the two analysis methods used, visualization was more successful. Among image categories used, images of faces stood out the best. The houses and scrambled images clustered to some extent as well. Visualizing electrodes resulted in pronounced clusters emerging, which were heterogeneous in their nature. Biclustering gave no noteworthy results.
This bachelor thesis explores how images are represented in human brain activity. Data used in this thesis were collected in the University Hospital of Lyon with experiments where human neural activity was recorded with intracranial electrodes during a simple visual task. The aim of this thesis is to analyse the data, more specifically the activity in the high-gamma frequency range, with unsupervised machine learning methods to find structure in it. In particular, similarity of neural responses (recorded by electrodes) to images and the differences in activity according to image categories.\n\rVisualization and biclustering were used in this thesis. Out of the two analysis methods used, visualization was more successful. Among image categories used, images of faces stood out the best. The houses and scrambled images clustered to some extent as well. Visualizing electrodes resulted in pronounced clusters emerging, which were heterogeneous in their nature. Biclustering gave no noteworthy results.