Reproduce “Learning multiple non-mutually-exclusive tasks for improved classification of inherently ordered labels” on OpenResearch

About this paper

Vadim Ratner, Yoel Shoshan, Tal Kachman

IBM Research introduces a Multi-Task Learning (MTL) framework that enhances classification performance for inherently ordered labels by creating multiple, non-mutually-exclusive binary tasks from the same dataset. This method demonstrated a nearly twofold increase in True Negative Rate at 0.95 True Positive Rate for challenging medical imaging classification tasks, effectively utilizing existing data without requiring new annotations.

Key insights

• Training on overlapping binary tasks implicitly conveys the inherent ordinal relationships between labels to the model, providing richer structural context.
• The distributed impact of potential label noise across multiple tasks enhances robustness against subjective label boundaries and annotator variability.
• Leveraging tasks with varying data balance allows the network to benefit from more balanced, 'easier' tasks to improve overall feature representation, which then aids performance on more challenging, imbalanced tasks, analogous to curriculum learning.

Summary by alphaXiv.