Towards Safe Weakly Supervised Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Weakly supervised learning (WSL) faces challenges with incomplete, inexact, or inaccurate labels.
• WSL performance can degrade with increased weakly supervised data, hindering real-world application.
• Developing safe WSL methods that guarantee performance is crucial.

Purpose of the Study:

• To propose a generic ensemble learning scheme for safe weakly supervised learning.
• To ensure predictions never seriously harm performance, even with limited or noisy labels.
• To address the deficiency of performance degradation in existing WSL methods.

Main Methods:

• A novel ensemble learning scheme integrating multiple weakly supervised learners.
• Optimization of worst-case performance gain using a maximin approach.
• Leveraging convex optimization (quadratic or linear programming) for efficient solutions.

Main Results:

• Guaranteed safe predictions for common convex loss functions under mild conditions.
• Flexible embedding of prior knowledge regarding base learner weights.
• Demonstrated effectiveness across diverse WSL tasks like semi-supervised learning, domain adaptation, multi-instance learning, and label noise learning.

Conclusions:

• The proposed ensemble method provides a robust solution for safe weakly supervised learning.
• The maximin optimization framework ensures reliable performance improvements.
• The approach is computationally efficient and offers intuitive geometric interpretations.