Could simplified stimuli change how the brain performs visual search tasks? A deep neural network study.

Area of Science:

• Computer Science
• Cognitive Science
• Neuroscience

Background:

• Visual search is complex, often studied with simplified stimuli that differ from natural images.
• This discrepancy may affect search behavior and be mistaken for cognitive processes like selective attention.

Purpose of the Study:

• To investigate how optimizing deep neural networks (DNNs) for one data distribution affects performance on a different distribution.
• To determine if data distribution differences explain performance limitations observed in visual search tasks.

Main Methods:

• Trained four DNN architectures on natural images, faces, or X-ray images.
• Adapted these DNNs to a visual search task using simplified stimuli.
• Compared DNN performance to models trained solely on the search task and to human performance.

Main Results:

• DNNs adapted to simplified stimuli showed human-like performance limitations.
• Models trained only on the search task did not exhibit these limitations.
• DNNs trained on natural images also showed limitations when adapted to a different natural image set.

Conclusions:

• Data distribution mismatch alone does not fully explain observed performance limitations.
• The findings suggest a need to reconsider how training data influences models of visual search.
• Future research should integrate optimization-based approaches into existing visual search models.