Evaluating (and Improving) the Correspondence Between Deep Neural Networks and Human Representations

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Machine Learning

Background:

• Psychological research has long sought to model human feature and category learning.
• Traditional models often rely on artificial stimuli, limiting real-world applicability.
• Deep neural networks (DNNs) now achieve human-level performance in image recognition.

Purpose of the Study:

• To investigate if DNNs can capture psychological representations of natural images.
• To identify discrepancies between DNNs and human similarity judgments.
• To improve DNN representations for predicting human category learning.

Main Methods:

• Utilized state-of-the-art object classification DNNs.
• Compared DNN similarity predictions with human judgments on natural images.
• Applied convex optimization to transform DNN representations.
• Used transformed representations to predict novel category learning difficulty.

Main Results:

• DNNs accurately predict human similarity judgments for natural images.
• DNNs fail to capture certain structures represented by humans.
• A convex optimization transformation significantly corrects these discrepancies.
• Improved DNN representations enhance predictions of category learning difficulty.

Conclusions:

• DNNs offer valuable insights into psychological representations of natural stimuli.
• A simple transformation can bridge the gap between DNNs and human cognition.
• This approach expands the use of natural stimuli in computational psychology.