Investigating Psychedelic Imagery through Convolutional Neural Networks’ Feature Visualization
Name
Carolin Lüübek
Abstract
Convolutional neural networks serve as primate visual system models and their feature visualization has intuitive similarities to psychedelic imagery, as well as the effect of features being similar in both artificial and biological systems. It is explored in this thesis whether the "psychedelicism" inherent in feature visualizations might signify deeper correspondences between convolutional neural networks and visual processing, particularly within the realm of psychedelic imagery.
The main research question of the thesis examines the variance in simulating psychedelic imagery between effective and ineffective visual system models. The results clarify that the feature visualizations of the CNN that approximates the visual system effectively are more accurate. This is speculated to point towards some of the computational mechanisms of effective visual system models being suitable for explaining the neural mechanisms underlying psychedelic imagery. Furthermore, the results hint at the heightened influence of endogenous activity from the primary visual area during psychedelic perception, as well as a pronounced alignment between artificial and biological systems at the neuronal level for early processing stages and at a more abstract level for later processing stages.
The main research question of the thesis examines the variance in simulating psychedelic imagery between effective and ineffective visual system models. The results clarify that the feature visualizations of the CNN that approximates the visual system effectively are more accurate. This is speculated to point towards some of the computational mechanisms of effective visual system models being suitable for explaining the neural mechanisms underlying psychedelic imagery. Furthermore, the results hint at the heightened influence of endogenous activity from the primary visual area during psychedelic perception, as well as a pronounced alignment between artificial and biological systems at the neuronal level for early processing stages and at a more abstract level for later processing stages.
Graduation Thesis language
English
Graduation Thesis type
Bachelor - Computer Science
Supervisor(s)
Jaan Aru
Defence year
2023