Year: 2022
Type: dissertation
Abstract: A call for deep neural networks (DNN) to permeate all classification problems has found its way to brain computer interface (BCI) applications. Despite this interest, the “depth” of DNNs in this field is a troublesome concept, and it appears uncon-vincing that the available data is being put to effective use. This dissertation centres around these questions by proposing ways to effectively treat the myriad recording contexts that are a necessary aspect of BCI applications. I argue that acknowledging these different contexts as domains of classification is an important step. Through doing so, I demonstrate that it is possible to develop deeper networks that make more effective use of available data as well as make use of the otherwise detrimental sensitivity DNNs exhibit to these domains. As a result of the latter, extensions are made into neuroscientific applications of DNNs. Finally, I argue for and propose an alternative path to pre-training extremely large DNNs inspired by language modelling rather than the extensions of successes in computer vision which have tended to be favoured. Finally, I describe the open source software library developed in support of this work, that enables similar work to continue in the future.
Source: TSpace (University of Toronto)
Author: Kostas, Demetres
Institution: University of Toronto
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7132985235