On Better Exploring and Exploiting Task Relationships in Multitask Learning: Joint Model and Feature Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence

Background:

• Multitask learning (MTL) leverages task interdependence for simultaneous learning.
• Existing MTL methods often independently learn shared parameters or features, leading to information loss.
• Measuring task relatedness is crucial for effective MTL.

Purpose of the Study:

• To propose a novel MTL strategy that jointly learns shared parameters and feature representations.
• To transform task features into a common space for enhanced task relatedness.
• To improve the optimization of shared parameters in MTL.

Main Methods:

• A new strategy for jointly learning shared parameters and feature representations in MTL.
• Feature transformation into a common feature space to increase task relatedness.
• An alternating algorithm for optimizing the non-convex objective function.

Main Results:

• A theoretical bound demonstrating improved task relatedness measurement.
• Experimental verification of the proposed joint model and feature MTL method's superiority.
• Demonstrated enhancement in optimizing shared parameters through joint learning.

Conclusions:

• The proposed MTL method effectively measures task relatedness by jointly learning shared parameters and features.
• This joint approach overcomes limitations of independent learning strategies.
• The method shows superior performance in various experimental settings.