The Comparison of Human and Machine Performance in Object Recognition.

Area of Science:

• Computer Vision
• Cognitive Science
• Artificial Intelligence

Background:

• Deep learning models claim human-level performance, but often under artificial conditions.
• Existing research shows mixed results on models aligning with human categorization behavior.
• Comparative psychology offers a framework for more rigorous human-AI comparisons.

Purpose of the Study:

• To investigate if deep learning models achieve human-level accuracy and human-like categorization.
• To compare model and human performance under similar constraints, using principles from comparative psychology.
• To explore model generalization, task adaptation, and qualitative categorization relationships.

Main Methods:

• Three experiments using subsets of the ObjectNet dataset.
• Varying presentation times, task complexities, and N-way vs. free-naming categorization.
• Analyzing object detection/classification separation and individual human differences.

Main Results:

• Models excel under machine-optimized conditions but struggle with real-world generalization without fine-tuning.
• Human categorization remains stable across tasks, while machine performance declines without adaptation.
• Only the multimodal CoCa model captured qualitative ordinal relationships in human categorization.

Conclusions:

• Current deep learning models approximate human categorization but do not fully replicate it, especially in complex, real-world scenarios.
• Comparative psychology principles and consideration of individual differences are crucial for evaluating AI.
• Further research is needed to bridge the gap between AI accuracy and human-like categorization.