Classification of faces in man and machine

Area of Science:

• Cognitive Science
• Computer Vision
• Machine Learning

Background:

• Understanding human face gender classification is crucial for AI development.
• Existing models often lack a deep understanding of human perceptual processes.

Purpose of the Study:

• To investigate the algorithms humans employ for face gender classification.
• To compare human classification strategies with various machine learning algorithms.
• To identify which machine learning models best replicate human decision-making in face perception.

Main Methods:

• Principal Component Analysis (PCA) was used on face image data.
• Machine learning classifiers (SVM, Relevance Vector Machine, Prototype Classifier, K-means) were trained on PCA eigenvectors and human gender estimates.
• Human and machine classification performance was analyzed by comparing errors, decision boundaries, and response metrics (reaction time, confidence).

Main Results:

• Classification error alone is insufficient to determine human-like algorithms.
• The distance of stimuli to the separating hyperplane (SH) effectively captures human internal decision spaces.
• Support Vector Machine (SVM) models, focusing on stimuli near the decision boundary, better represent human classification than prototype classifiers.

Conclusions:

• Human face gender classification is better modeled by algorithms sensitive to decision boundaries.
• Machine learning, particularly SVM, can provide insights into human cognitive processes for image perception.
• Future AI systems could benefit from incorporating human-like decision boundary sensitivity for improved face analysis.