What question did this study set out to answer?

This research aims to quantify individual differences in aesthetic cognition using a new computational framework.

April 24, 2026

Does Beauty Follow Patterns? — A Computational Mechanism for Aesthetic Cognition Based on the CLIP Model

NWNa WANGQingdao National Laboratory for Marine Science and Technology

Key Points

This research aims to quantify individual differences in aesthetic cognition using a new computational framework.
Constructed a cross-modal aesthetic cognitive computing framework based on the CLIP model.
Used StyleGAN2 to embed aesthetic style features in the visual branch.
Utilized a custom dataset and implemented cognitive consistency loss to optimize the model's performance.
Achieved 92.7% accuracy in aesthetic style classification.
Established a Pearson correlation coefficient of 0.89 with human evaluation scores (p<0.001).
Enabled numerical analysis and anomaly detection in aesthetic perception.

Abstract

Aesthetic cognition reflects both artistic experience and underlying neuropsychological state and cognitive function. Quantifying individual perceptual differences in complex visual stimuli provides a non-invasive engineering path for mental and psychological rehabilitation and cognitive impairment screening. To address the limitations of subjective and static cognitive assessments, this study constructs a cross-modal aesthetic cognitive computing framework based on generative artificial intelligence (AI). This framework is based on the Contrastive Language-Image Pre-training (CLIP) model, embedding structured aesthetic style features generated by Style-Based Generative Adversarial Network 2 (StyleGAN2) in the visual branch. Its latent spatial characteristics are utilized as a "gene map" of visual stimuli to enhance the representation of formal patterns. The text branch utilizes Bidirectional Encoder Representation from Transformers (BERT) based on Transformers to construct a hierarchical semantic distillation module, transforming abstract aesthetic concepts into computable vectors. By introducing a cognitive consistency loss function, the model jointly optimized the image text style embedding space, making its output highly fit the trajectory of human aesthetic cognition. Based on a self-built Surreal Symphonies dataset containing 30 genres and human cognitive behavior data, experiments showed that the accuracy of the enhanced model's aesthetic style classification reached 92.7%, and the Pearson correlation coefficient with human evaluation scores was 0.89 (p<0.001). This method effectively realizes the numerical deconstruction and anomaly detection of aesthetic perception. This study provides a potential generative artificial intelligence–based tool and methodological reference for quantifying cognitive differences. This strategy advances personalized art research and supports future applications in mental health assessment and cognitive function research.

Ask AI

Helpful

Bookmark

View Full Paper