Making a Bird AI Expert Work for You and Me

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction
• Computer Vision

Background:

• Fine-grained visual classification (FGVC) models excel at distinguishing subtle differences in images, but their outputs (e.g., specific bird species) may not directly translate to human learning.
• A key challenge lies in determining what constitutes transferable knowledge from AI to humans and how to effectively measure this knowledge gain.

Purpose of the Study:

• To investigate how AI can serve as a knowledge provider to enhance human expertise in specific domains.
• To identify the most effective forms of transferable knowledge from AI models and develop practical methods for measuring human expertise improvement.

Main Methods:

• Proposed representing AI knowledge as highly discriminative, expert-exclusive visual regions.
• Developed a novel multi-stage learning framework to instantiate this knowledge transfer.
• Conducted a large-scale human study involving 15,000 trials to evaluate the method's effectiveness.

Main Results:

• The proposed method significantly improved participants' ability to recognize previously unidentifiable bird species.
• The AI-driven knowledge transfer was effective across individuals with varying levels of prior expertise.
• A new metric, TEMI (Transferable Expertise Measurement Instrument), was proposed for benchmarking future research.

Conclusions:

• AI can be a valuable tool for transferring specialized knowledge to humans, enhancing their domain expertise.
• Focusing on discriminative visual regions as knowledge representations is a promising approach for AI-to-human learning.
• The developed framework and TEMI metric provide a foundation for future research in human-AI collaborative learning.