Color illusions also deceive CNNs for low-level vision tasks: Analysis and implications.

Area of Science:

• Computational neuroscience
• Computer vision
• Psychophysics

Background:

• Visual illusions offer insights into human vision science.
• Convolutional neural networks (CNNs) are increasingly used to model visual processing.

Purpose of the Study:

• To investigate similarities and differences between CNNs and human perception regarding visual illusions.
• To analyze the origins of these similarities and differences in CNN architectures.

Main Methods:

• Training CNNs on low-level visual tasks.
• Applying linear eigenanalysis to simple CNNs.
• Comparing CNN-generated illusions with human perception data.

Main Results:

• Simple CNNs, when linearized, exhibit properties similar to human visual systems (e.g., center-surround receptive fields, contrast sensitivity functions).
• Complex, nonlinear CNN architectures may fail to replicate human visual illusions accurately.
• Differences in illusion replication stem from nonlinear behaviors not captured by linear approximations.

Conclusions:

• Low-level visual illusions in simple CNNs may arise from optimization processes, mirroring human vision.
• Complex CNNs, despite their flexibility, might not be ideal for studying human vision due to diverging nonlinear dynamics.
• Caution is advised when using CNNs to model human vision, considering inherent limitations and architectural nonlinearities.