AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-22
Model-agnostic information transfer and fusion for classification with label noise
The researchers developed a model-agnostic, nonparametric framework for classification using a large dataset with noisy labels and a small, expert-verified clean dataset. It uses the clean data to purify the noisy dataset while carefully handling samples that remain ambiguous. The framework has statistical theory behind it, and its performance was demonstrated in simulations and medical image analysis for pneumonia diagnosis.