Read the original document by opening this link in a new tab.
Table of Contents
1 Introduction
2 DAL Approaches
3 Comparative Experiments of DAL
Summary
Deep Active Learning (DAL) has emerged as a solution to maximize model performance under limited labeling costs. This paper presents a survey of DAL methods, constructing a DAL toolkit, DeepAL+, and conducting comparative experiments. The work explores factors influencing DAL efficacy and offers guidelines for researchers. Various querying strategies in DAL, such as uncertainty-based, diversity-based, and combined strategies, are discussed. Enhancement methods for DAL include data augmentation, pseudo labeling, and model improvements through extra networks and ensemble learning. Comparative experiments on 19 DAL methods across multiple datasets provide insights into the performance and challenges of DAL.