References & Citations
Quantitative Biology > Neurons and Cognition
Title: Contrast Sensitivity Functions in Autoencoders
(Submitted on 28 Feb 2021 (v1), revised 12 Mar 2021 (this version, v2), latest version 4 Mar 2022 (v6))
Abstract: Artificial neural networks (ANNs) trained to solve low-level visual tasks have been shown to develop biological-like features. Examples include human-like receptive fields and being subject to human-like visual illusions. In this work we report that these artificial systems may also develop human-like Contrast Sensitivity Functions (CSFs).
Following the interest in accurate comparison between humans and ANNs, we use psychophysical ideas to extend a recently proposed eigenanalysis to define the CSFs of such artificial systems. Results show that low-pass sensitivity to chromatic gratings and wider band-pass sensitivity to achromatic gratings may appear in convolutional autoencoders trained to enhance the retinal signals. Similarly, 3D versions of these networks may develop a spatio-temporal CSF with the basic diamond shape and bandwidth of the human window of visibility. Moreover, masking-like saturated responses to the contrast of spatio-chromatic and spatio-temporal gratings may emerge in these autoencoders.
However, the consideration of a range of ANN models trained for the same statistical goal shows that deeper architectures or better accuracy in the goal do not necessarily lead to more human CSFs or more human masking-like saturations. We argue that the nontrivial interplay between goal and architecture in the optimization prevents against premature conclusions about the organization principles of the visual system if sensible architectures are not adopted. This suggests that goal-driven derivations of psychophysical phenomena should include more realistic units than those considered in conventional ANNs.
Submission history
From: Jesus Malo [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)
[v5] Sun, 5 Dec 2021 14:04:44 GMT (11823kb,D)
[v6] Fri, 4 Mar 2022 05:10:39 GMT (8373kb,D)
Link back to: arXiv, form interface, contact.