Efficient Task Grouping Through Sample-Wise Optimisation Landscape Analysis.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Optimization

Background:

• Shared training methods like multi-task learning (MTL) can cause negative transfer, degrading performance.
• Identifying optimal task combinations for joint learning (task grouping) is computationally challenging due to combinatorial explosion.

Purpose of the Study:

• To develop an efficient framework for task grouping to mitigate computational challenges in shared training.
• To reduce the need for extensive model training and evaluation cycles in task combination selection.

Main Methods:

• Infers pairwise task similarities using sample-wise optimization landscape analysis, avoiding shared model training.
• Employs a graph-based clustering algorithm to identify near-optimal task groups for efficient joint learning.

Main Results:

• Achieves a five-fold speed enhancement compared to state-of-the-art methods across 9 diverse datasets.
• Demonstrates comparable performance to existing methods, validating the framework's efficiency and effectiveness.

Conclusions:

• The proposed framework offers an efficient and effective solution for the NP-hard task grouping problem.
• Enables faster and more practical application of shared training techniques by optimizing task combinations.