Crowding reveals fundamental differences in local vs. global processing in humans and machines

Area of Science:

• Computational neuroscience
• Computer vision
• Cognitive science

Background:

• Feedforward Convolutional Neural Networks (ffCNNs) are advanced models in computer vision and neuroscience.
• Human-like performance in ffCNNs does not guarantee human-like computational processes.
• Prior research indicates ffCNNs may not utilize global shape information.

Purpose of the Study:

• To investigate whether ffCNNs can perform human-like global shape computations.
• To determine if limitations in ffCNNs stem from architecture or training.
• To use visual crowding as a probe for global shape processing.

Main Methods:

• Employing visual crowding as a precise experimental paradigm.
• Analyzing the computational capabilities of ffCNNs in processing global shape information.
• Comparing ffCNN performance against human visual processing benchmarks.

Main Results:

• ffCNNs demonstrate an inability to perform human-like global shape computations.
• Evidence suggests architectural constraints are the primary reason for this limitation.
• The findings highlight a fundamental difference in how ffCNNs and humans process visual information.

Conclusions:

• ffCNNs possess inherent architectural limitations preventing human-like global shape processing.
• Future models require architectural modifications to better emulate the human visual system.
• Addressing these shortcomings is crucial for developing more accurate computational models of human vision.