Task Similarity Estimation Through Adversarial Multitask Neural Network.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Multitask learning (MTL) leverages shared knowledge across related tasks to enhance individual task performance.
• Theoretical understanding of task relationships and their impact on shared knowledge in MTL remains an open challenge.

Purpose of the Study:

• To develop a theoretical perspective on the benefits of using information similarity in MTL.
• To propose practical principles for controlling generalization errors using similarity information.
• To introduce a novel training algorithm for multitask neural networks.

Main Methods:

• Proposed an upper bound on generalization error using Wasserstein distance as a similarity metric.
• Developed a new training algorithm for adversarial multitask neural networks.
• Learned task relation coefficients and neural network parameters automatically.

Main Results:

• Demonstrated improved empirical performance on computer vision benchmarks.
• Showcased the advantage of the proposed approach in extracting task relations from real medical datasets.
• Validated the practical principles for controlling generalization errors in MTL.

Conclusions:

• The theoretical framework provides insights into leveraging information similarity for effective MTL.
• The novel training algorithm enhances MTL performance and task relation extraction.
• The approach shows promise for applications in computer vision and medical data analysis.