Mapping of convolutional neural network activation maps on the visual cortex using a Bayesian linear framework
Keywords
Loading...
Authors
Issue Date
2016-03-29
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Decoding the brain is slowly starting to become a reality. Recently, new computational models
have made it possible to make reconstructions of perceived images from BOLD responses in the
visual cortex. A linear Gaussian framework was used by a previous study to decode functional
magnetic resonance imaging data from subjects who were presented with images of handwritten
characters. Here we expand this research by taking convolutional neural network activation maps
instead of images of handwritten characters as the encoding and decoding stimuli as the di erent
layers in the network appear to be good models for the di erent layers in the visual cortex. In
contrast to former studies, a uniform prior is used to let the decoding be driven by the found
encoding only. This approach results in good classi cation accuracies that become better as the
convolutional layer becomes deeper and are more robust than previously found results.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen