Effects of Biases in Geometric and Physics-Based Imaging Attributes on Classification Performance.

Area of Science:

• Computer Vision
• Machine Learning
• Cognitive Science

Background:

• Learned systems in visual recognition generalize well despite small training datasets.
• Small datasets risk containing biases, impacting system performance on new data.
• Selection bias, specifically in image data collection, is a significant concern.

Purpose of the Study:

• To investigate the limits of generalization in learned visual systems when faced with selection bias.
• To understand how biases in imaging physics and geometry affect machine learning.
• To theoretically and empirically evaluate the impact of selection bias on classification.

Main Methods:

• Theoretical analysis using thought experiments to probe data collection biases.
• Development of theoretical tools to identify deficiencies in data sampling.
• Empirical testing on a new dataset using established top classifiers.

Main Results:

• Theoretical results highlight potential deficiencies in data collection and system development.
• Empirical tests confirm that certain selection biases challenge existing classifiers.
• Biases rooted in the physics and imaging geometry of vision were identified as problematic.

Conclusions:

• Selection bias in visual data, influenced by imaging physics and geometry, poses a challenge to current machine learning classification.
• Both theoretical and empirical findings underscore the need for careful data collection and bias mitigation strategies.
• Further research is needed to address these vision-specific biases for more robust AI systems.