On Statistical Bias In Active Learning: How and When To Fix It

2021/01/27 by Sebastian Farquhar, Yarin Gal, Farquhar, Sebastian +3 · 24 citations
Computer Science · Engineering · Mathematics · #Active learning (machine learning) #Algorithms and Data Compression #Artificial intelligence #Computer science #Engineering #FOS: Computer and information sciences #Inductive bias #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Machine Learning and Algorithms #Machine Learning and Data Classification #Machine learning #Mechanism (biology) #Multi-task learning #Population #Statistical learning #cs.LG #stat.ML

paper · pdf · doi:10.48550/arxiv.2101.11665

published in arXiv (Cornell University) (Cornell University) · Published at ICLR 2021 (Spotlight)

openalex publication_date 2021/01/27 · arxiv created 2021/05/31 · arxiv updated 2021/06/01 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Active learning is a powerful tool when labelling data is expensive, but it introduces a bias because the training data no longer follows the population distribution. We formalize this bias and investigate the situations in which it can be harmful and sometimes even helpful. We further introduce novel corrective weights to remove bias when doing so is beneficial. Through this, our work not only provides a useful mechanism that can improve the active learning approach, but also an explanation of the empirical successes of various existing approaches which ignore this bias. In particular, we show that this bias can be actively helpful when training overparameterized models -- like neural networks -- with relatively little data.

Citations

• Bayesian Active Learning for Classification and Preference Learning
• Information-Based Objective Functions for Active Data Selection
• On the Convergence of Adam and Beyond
• Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning

Cited by

• ShaplEIG: Bayesian Experimental Design for Shapley Value Estimation
• Scaling Up Active Testing to Large Language Models
• Prediction-Oriented Subsampling from Data Streams
• Consensus-Driven Active Model Selection
• Step-DAD: Semi-Amortized Policy-Based Bayesian Experimental Design
• Active Measurement: Efficient Estimation at Scale
• Representative, Informative, and De-Amplifying: Requirements for Robust Bayesian Active Learning under Model Misspecification
• Active Learning Methods for Efficient Data Utilization and Model Performance Enhancement

Related

• Active Bias: Training More Accurate Neural Networks by Emphasizing High Variance Samples
• Addressing Bias in Active Learning with Depth Uncertainty Networks... or Not
• Overcoming Overconfidence for Active Learning
• Towards Robust and Reproducible Active Learning Using Neural Networks
• Not All are Made Equal: Consistency of Weighted Averaging Estimators Under Active Learning