The application of explainable artificial intelligence methods to models for automatic creativity assessment

Area of Science:

• Psychology
• Computer Science
• Artificial Intelligence

Background:

• Assessing creativity is crucial for understanding human cognition.
• The Urban test provides a standardized method for evaluating creativity through drawings.
• Automated assessment of creativity can offer objective and scalable insights.

Purpose of the Study:

• To compare Artificial Intelligence (AI) models for creativity level determination using Urban test drawings.
• To apply explainable AI (XAI) methods to identify key drawing features influencing AI predictions.
• To validate AI-driven creativity assessment against expert evaluations.

Main Methods:

• Utilized a dataset of 1,823 Urban test drawings.
• Fine-tuned pre-trained AI models (MobileNet, ResNet18, AlexNet, DenseNet, ResNext, EfficientNet, ViT) for score prediction.
• Employed Gradient-weighted Class Activation Mapping (Grad-CAM) for visualizing model attention areas.

Main Results:

• MobileNet achieved the highest prediction accuracy at 76%.
• XAI analysis revealed that model-identified features correlated with expert Urban test criteria.
• Analysis of incorrect predictions helped refine understanding of AI reliance on specific drawing features.

Conclusions:

• Neural network methods can automate creativity level diagnosis via the Urban test.
• XAI confirmed that identified AI activation zones align with expert Urban test assessment rules.
• This approach offers a promising tool for objective creativity evaluation.