Contrast sensitivity functions in autoencoders.

Area of Science:

• Computational neuroscience
• Computer vision
• Human visual perception

Background:

• Human contrast sensitivity functions (CSFs) are fundamental to visual perception.
• Previous theories suggested CSFs arise from efficient retinal image analysis.
• Artificial neural networks (ANNs) are increasingly used to model visual processing.

Purpose of the Study:

• To reassess the role of low-level vision tasks in explaining human CSFs.
• To investigate the utility of convolutional neural networks (CNNs) in modeling CSFs.
• To explore the relationship between network depth, task optimization, and human-like visual phenomena.

Main Methods:

• Trained autoencoder CNNs on various low-level image processing tasks (e.g., noise/blur removal).
• Evaluated the ability of trained CNNs to replicate human CSFs across spatiotemporal and chromatic dimensions.
• Compared the performance of different CNN architectures and task objectives.

Main Results:

• Autoencoders trained on noise and blur removal tasks successfully developed human-like CSFs.
• The best-performing CNN model reproduced CSFs with an 11% root mean square error.
• Deeper CNNs optimized for quantitative goals showed reduced ability to replicate human-like CSFs.

Conclusions:

• Low-level vision tasks, when modeled by appropriate CNNs, can explain human CSFs.
• Network architecture and task optimization goals are critical for modeling human vision accurately.
• Caution is advised when using ANNs in vision science due to potential limitations in simplified units or architectures.