Robust Multi-Task Learning With Flexible Manifold Constraint.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multi-task learning (MTL) leverages related tasks to improve model performance.
• Existing MTL methods often degrade significantly in the presence of data outliers.
• Robustness to noisy data is crucial for real-world applications of MTL.

Purpose of the Study:

• To develop a novel robust multi-task learning model capable of handling polluted data.
• To enhance the performance and reliability of multi-task learning in the presence of outliers.
• To introduce a flexible manifold constraint and a robust loss function for improved MTL.

Main Methods:

• Proposed a novel robust multi-task model named FMC-MTL.
• Incorporated a flexible manifold constraint (FMC) using a relaxed and generalized Stiefel Manifold.
• Developed a robust loss function interpolating between l2,1-norm and squared Frobenius norm.

Main Results:

• FMC-MTL demonstrates remarkable robustness to contaminated data.
• The model effectively handles severely polluted datasets.
• Experimental results show superiority over state-of-the-art multi-task models on noisy data.

Conclusions:

• FMC-MTL offers an effective solution for multi-task learning with outlier-polluted data.
• The proposed approach preserves data structure while ensuring robustness.
• This work advances the field of robust multi-task learning.