Title: Weakly Supervised and Noisy Label Learning: When Ground Truth Is Imperfect
Authors: Virendra Tank
Volume: 10
Issue: 6
Pages: 137-144
Publication Date: 2026/06/28
Abstract:
The perfect, ideal clean labelled data assumption does not hold in actual machine learning problems. Here, we survey several techniques for learning from imperfect supervision such as missing annotation, noisy label and weak supervision cues. We review multi-instance learning frameworks, junk label de-noising and outlier removal methods, week supervision based object detection & segmentation algorithms as well as crowd-based label aggregation schemes. By dissecting the theory and phenomena learned with modern deep learning on theories in medical imaging and web-scale data, we establish that reliable knowledge can be extracted autonomously from unreliable supervision. This review presents recent progress in weakly supervised learning and discusses the main challenges and future perspectives for learning with imperfect ground truth.