Differential Involvement of EEG Oscillatory Components in Sameness versus Spatial-Relation Visual Reasoning Tasks

Area of Science:

• Cognitive Neuroscience
• Computer Vision
• Computational Neuroscience

Background:

• Deep convolutional neural networks (CNNs) are successful computational models for computer vision.
• However, CNNs show limitations in tasks beyond simple image categorization.
• The same-different (SD) judgment task highlights these limitations compared to spatial relationship (SR) tasks.

Purpose of the Study:

• To investigate the limitations of current computational models, specifically CNNs, in visual cognition.
• To test if same-different (SD) tasks recruit additional cortical mechanisms not captured by CNNs.
• To compare human neural activity with CNN performance on SD and SR tasks.

Main Methods:

• Recorded electroencephalography (EEG) signals from human participants performing SD and SR tasks.
• Matched task difficulty for human subjects using an adaptive psychometric procedure.
• Compared human EEG data with the performance of artificial neural networks on the same tasks.

Main Results:

• Human participants showed higher activity in the low beta (16-24 Hz) frequency band during SD tasks compared to SR tasks.
• Evoked potentials (EPs) were modulated differently between the two tasks.
• These findings suggest SD tasks engage additional neural mechanisms beyond those in current CNNs.

Conclusions:

• Current feed-forward CNNs likely lack essential mechanisms like working memory and attention.
• These additional mechanisms are crucial for complex visual cognition, particularly for same-different judgments.
• Future computational models need to incorporate these mechanisms to better replicate human visual cognition.