Introducing Local Weight Autocorrelation in Deep Neural Networks leads to the Emergence of Orientation Columns and Face Patches
Keywords
Loading...
Authors
Issue Date
2021-01-04
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Smoothly varying orientation preference maps and category-selective regions are well-known prop-
erties of the primate visual system. With the aim of reproducing these properties in arti cial neural
network models of the ventral stream, we introduce the Hypercolumn layer, which is a subclass of the
locally-connected layer and an abstraction of the in
uential cortical Hypercolumn model. Our Hyper-
column layer minimizes the spatial distance of the layer weight vectors, which by de nition optimizes
a local weight autocorrelation. We show that deep neural networks using Hypercolumn layers produce
smooth orientation preference maps in shallow layers and category-selective regions in deeper layers when
trained on a categorical classi cation task. The strength of category selectivity is proportionate to the
degree of visual expertise that the model has with the category. We nd no substantial accuracy bene ts
of optimizing weight autocorrelation, although it reduces over tting. The biological implications of the
models are thoroughly discussed.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen