We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

q-bio.NC

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Quantitative Biology > Neurons and Cognition

Title: Contrast Sensitivity Functions in Autoencoders

Abstract: Three decades ago, Atick et al. suggested that human frequency sensitivity may emerge from the enhancement required for a more efficient analysis of retinal images. Here we reassess the relevance of low-level vision tasks in the explanation of the Contrast Sensitivity Functions (CSFs) in light of (1) the current trend of using artificial neural networks for studying vision, and (2) the current knowledge of retinal image representations.
As a first contribution, we show that a very popular type of convolutional neural networks (CNNs), called autoencoders, may develop human-like CSFs in the spatio-temporal and chromatic dimensions when trained to perform some basic low-level vision tasks (like retinal noise and optical blur removal), but not others (like chromatic adaptation).
As a second contribution, we provide experimental evidence of the fact that, for some functional goals (at low abstraction level), deeper CNNs that are better in reaching the quantitative goal are actually worse in replicating human-like phenomena (such as the CSFs). This low-level result is not necessarily in contradiction with other works that report advantages of deeper nets in modeling higher-level vision goals. However, in line with a growing body of literature, our results suggests another word of caution about CNNs since the use of simplified units or unrealistic architectures in goal optimization may be a limitation for the modeling and understanding of human vision.
Subjects: Neurons and Cognition (q-bio.NC)
Cite as: arXiv:2103.00481 [q-bio.NC]
  (or arXiv:2103.00481v4 [q-bio.NC] for this version)

Submission history

From: Qiang Li [view email]
[v1] Sun, 28 Feb 2021 12:19:02 GMT (3773kb,D)
[v2] Fri, 12 Mar 2021 14:23:41 GMT (3774kb,D)
[v3] Thu, 12 Aug 2021 13:49:27 GMT (5803kb,D)
[v4] Fri, 13 Aug 2021 13:49:35 GMT (5801kb,D)

Link back to: arXiv, form interface, contact.