ViSpa (Vision Spaces): A computer-vision-based representation system for individual images and concept prototypes,...

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Computer Vision

Background:

• Quantitative models of mental representations are established in language but lacking for vision.
• Previous models did not capture vision-specific similarities for naturalistic images.

Purpose of the Study:

• Introduce ViSpa (Vision Spaces), a novel computational model for vision-based mental representations.
• Validate ViSpa's ability to predict human behavioral data across diverse tasks.
• Provide a tool for researchers to control experimental materials and analyze visual similarity.

Main Methods:

• Developed ViSpa using a deep convolutional neural network trained on image classification.
• Generated high-dimensional vector spaces representing naturalistic images and concept prototypes.
• Evaluated ViSpa's vision-based similarity scores against human behavioral data in five large-scale studies (similarity, typicality, discrimination, priming).

Main Results:

• ViSpa similarity scores significantly predicted human behavioral data across all tested tasks.
• Demonstrated ViSpa's effectiveness in modeling visual similarity for both image pairs and word pairs referring to imageable concepts.
• Confirmed ViSpa's utility in predicting online processing times and error rates.

Conclusions:

• ViSpa provides a theoretically grounded and empirically validated model for vision-based mental representations.
• ViSpa serves as a valuable tool for psychological research, enabling precise control over experimental stimuli and data analysis.
• The accessibility of ViSpa through tutorials and a web interface promotes its adoption in the research community.