Unprejudiced Training Auxiliary Tasks Makes Primary Better: A Multitask Learning Perspective.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Multitask learning (MTL) leverages knowledge from related tasks to improve primary task performance.
• Current MTL methods often under-train auxiliary tasks by assigning them lower loss weights than the primary task.
• This imbalance limits the effectiveness of auxiliary tasks in supporting the primary objective.

Purpose of the Study:

• To propose an uncertainty-based impartial learning method for balanced multitask training.
• To enhance neural network performance on a primary task by effectively utilizing auxiliary tasks.
• To develop a robust weighting strategy for auxiliary losses that accounts for task uncertainty.

Main Methods:

• Implemented an uncertainty-based impartial learning approach to ensure balanced task training.
• Incorporated both gradient and uncertainty information during backpropagation for improved primary task focus.
• Developed a novel weighting strategy for auxiliary losses that dynamically adjusts based on task uncertainty.

Main Results:

• The proposed method achieved performance comparable to or exceeding state-of-the-art MTL approaches.
• Demonstrated effective and robust enhancement of primary task performance.
• Showed that the weighting strategy is resilient to noise in auxiliary task pseudolabels.

Conclusions:

• Uncertainty-based impartial learning provides a balanced and effective approach to multitask learning.
• Considering task uncertainty and gradients during training significantly boosts primary task performance.
• The proposed method offers a robust solution for leveraging auxiliary tasks, even with noisy data.