Dual-balancing for multi-task learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Deep Learning

Background:

• Multi-task learning (MTL) enables learning multiple related tasks simultaneously, showing success across various domains.
• A key challenge in MTL is performance compromise due to disparities in task loss and gradient scales.
• Effective task balancing remains crucial for optimizing MTL performance.

Purpose of the Study:

• To introduce Dual-Balancing Multi-Task Learning (DB-MTL) for effective task balancing.
• To address performance compromises arising from imbalanced loss and gradient scales in MTL.
• To improve overall performance in multi-task learning scenarios.

Main Methods:

• DB-MTL balances task losses using logarithm transformation.
• Gradient magnitudes are rescaled via normalization to comparable magnitudes using the maximum gradient norm.
• The proposed method integrates both loss-scale and gradient-scale balancing strategies.

Main Results:

• DB-MTL demonstrates consistent performance improvements across multiple benchmark datasets.
• The proposed method effectively mitigates the negative impact of task imbalance.
• Experimental results show DB-MTL outperforms current state-of-the-art methods.

Conclusions:

• DB-MTL offers a robust solution for task balancing in multi-task learning.
• The dual-balancing approach enhances model performance by addressing both loss and gradient disparities.
• DB-MTL represents a significant advancement in optimizing multi-task learning effectiveness.