Slice-based Learning: A Programming Model for Residual Learning in Critical Data Slices.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Machine learning models often underperform on critical data subsets, termed slices, despite high overall accuracy.
• Existing methods like training separate expert models or multi-task learning have limitations in addressing slice-level performance.

Purpose of the Study:

• To introduce Slice-based Learning, a novel programming model to improve performance on critical data slices.
• To enable models to commit additional capacity to specified critical data subsets using slicing functions (SFs).

Main Methods:

• Developed a programming model utilizing slicing functions (SFs) to define critical data subsets (slices).
• Proposed learning slice expert representations combined with an attention mechanism for slice-aware predictions.
• Applied the approach to diverse datasets including language understanding, computer vision, and industrial systems.

Main Results:

• Achieved significant improvements over baseline methods, with up to 19.0 F1-score gains on slices.
• Demonstrated an overall F1-score improvement of up to 4.6 points across various datasets.
• Maintained a parameter-efficient representation while enhancing performance on critical data subsets.

Conclusions:

• Slice-based Learning offers an effective and parameter-efficient method for improving machine learning model performance on critical data slices.
• The proposed approach enhances model awareness of and performance on specific, important data subsets, leading to better real-world applicability.