Learning from Ambiguously Labeled Face Images.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Supervised learning requires accurately labeled data, which is often unavailable for real-world datasets like news photos.
• Ambiguously labeled face images pose a significant challenge for training effective classifiers.
• Existing methods struggle with label imbalance and iterative refinement in ambiguous scenarios.

Purpose of the Study:

• To develop a robust method for resolving ambiguity in face image labels.
• To improve the accuracy of classifiers trained on ambiguously labeled data.
• To address label imbalance and enable reliable iterative refinement of predictions.

Main Methods:

• Proposed Matrix Completion for Ambiguity Resolution (MCar) to predict actual labels from ambiguous images.
• Introduced weighted MCar (WMCar) to handle label imbalance.
• Developed Iterative Candidate Elimination (ICE) for stable iterative refinement and incorporated labeling constraints.

Main Results:

• MCar and WMCar effectively predict labels from ambiguously labeled face images.
• The ICE procedure enables reliable iterative refinement, overcoming performance degradation from noisy labels.
• Extended MCar successfully incorporates prior labeling constraints.
• Demonstrated significant improvements over existing methods on multiple ambiguously labeled datasets.

Conclusions:

• The proposed MCar, WMCar, and ICE methods offer a powerful solution for training classifiers with ambiguously labeled face image data.
• The approach effectively handles label imbalance and enables stable iterative learning.
• This work advances the field of machine learning for datasets with inherent labeling uncertainties.