Perceptual Dominance in Brief Presentations of Mixed Images: Human Perception vs. Deep Neural Networks

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Visual perception involves selecting salient stimuli and suppressing others.
• Deep neural networks (DNNs) show promise as models for the brain's ventral visual stream due to representational similarities.
• Understanding visual competition is key to understanding perceptual decision-making.

Purpose of the Study:

• To investigate visual competition in humans and feed-forward DNNs using brief image mixtures.
• To identify image parameters influencing perceptual dominance in both systems.
• To compare the predictive power of these parameters across human and DNN visual systems.

Main Methods:

• Created non-dynamic visual competition tasks by briefly presenting human participants with dual-image mixtures.
• Tested feed-forward DNNs with identical image mixtures to observe their responses.
• Analyzed image parameters to determine their influence on which image was perceived.

Main Results:

• Both humans and DNNs predominantly perceived only one image from the presented mixtures.
• Identified specific image parameters that predict perceptual dominance.
• Demonstrated comparable predictability of these parameters for human and DNN visual processing.

Conclusions:

• Human and DNN visual systems exhibit analogous behaviors in perceptual competition.
• Image characteristics significantly influence visual dominance in both biological and artificial systems.
• Findings can enhance DNNs as models of the brain and improve DNN performance by incorporating biological insights.