LiNC: Lightweight Noise Correction via Adaptive Label Refinement

Abhishek Moturu, Babak Taati, Anna Goldenberg

Keywords: noise, correction, trust, lightweight, mislabels

TL;DR: assign each sample a trainable trust parameter that shifts from label to prediction to correct noise

Abstract: Medical imaging datasets often contain label noise due to factors such as inter-rater variability, annotation errors, and ambiguous cases, which may severely undermine the reliability and clinical effectiveness of machine learning models trained using those datasets. To address this challenge, we introduce Lightweight Noise Correction (LiNC), which is an intuitive and powerful approach that assigns a trainable trust parameter, $\alpha_i$, to each individual training sample. Initially initialized to fully trust the observed labels, these parameters adaptively shift trust towards model predictions through a gradient-based optimization process, effectively identifying and reducing the impact of noisy labels by correcting them. After this correction process, usual model training is carried out. Our method requires minimal computational overhead, making it practical for widespread adoption in cases where noise is suspected within a dataset. Extensive evaluations on ten medical imaging datasets from the MedMNISTv2 collection reveal significant improvements in classification accuracy and AUROC across various uniform label noise levels (ranging from 0\% to 50\%) and robust detection of mislabeled samples, underscoring LiNC's potential to improve noisy machine learning.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 5426