Deep learning models fail to capture the configural nature of human shape perception

Area of Science:

• Cognitive Science
• Neuroscience
• Computer Vision

Background:

• Human object perception relies on holistic configuration of local shape features.
• Deep convolutional neural networks (DCNNs) are leading models for visual object recognition.
• It remains unclear if DCNNs replicate human configural sensitivity.

Purpose of the Study:

• To investigate whether DCNNs exhibit configural sensitivity in object recognition.
• To compare DCNN performance with human performance on a configural task.
• To explore modifications for enhancing DCNNs' brain-like processing.

Main Methods:

• Utilized a dataset of animal silhouettes.
• Created a modified dataset disrupting object configuration while preserving local features.
• Assessed human and DCNN performance on both original and modified datasets.

Main Results:

• Human performance significantly decreased with configuration disruption.
• DCNN performance remained unaffected by the configuration disruption.
• Network modifications and training strategies did not induce configural processing.
• DCNNs failed to predict human trial-by-trial object judgments.

Conclusions:

• Current DCNNs are insensitive to object configuration, unlike humans.
• Achieving human-like configural sensitivity may require training DCNNs on broader object-related tasks beyond simple category recognition.