Extracting Low-Dimensional Psychological Representations from Convolutional Neural Networks.

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Computer Vision

Background:

• Convolutional neural networks (CNNs) are prevalent in psychology and neuroscience for predicting human responses to visual stimuli.
• CNNs typically utilize thousands of learned features for image representation, raising questions about feature necessity for modeling human behavior.

Purpose of the Study:

• To estimate the number of dimensions required in CNN representations to accurately capture human psychological representations.
• To investigate the interpretability of reduced-dimension CNN features for understanding visual information processing.

Main Methods:

• Direct estimation using human similarity judgments.
• Indirect estimation within a categorization task.
• Analysis of low-dimensional projections of CNN representations.

Main Results:

• Low-dimensional projections of CNN representations were sufficient for predicting human behavior in both similarity judgments and categorization tasks.
• These reduced representations were interpretable, offering insights into human visual information processing.
• Control studies confirmed findings were robust and not dataset-size dependent, suggesting feature redundancy in CNNs.

Conclusions:

• A smaller subset of features from CNNs can effectively model human psychological and neural representations of visual stimuli.
• The interpretability of these low-dimensional features facilitates understanding the mechanisms of human visual perception.
• Findings suggest significant redundancy in standard CNN feature sets for modeling human visual cognition.