# A Comparative Survey of Deep Active Learning

**Authors:** Xueying Zhan, Qingzhong Wang, Kuan-hao Huang, Haoyi Xiong, Dejing Dou, Antoni B. Chan  
**arXiv:** [2203.13450](https://arxiv.org/abs/2203.13450)  
**Date:** 2022-03-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2203.13450v3/

## Key points

- Deep learning relies on extensive labeled data for good performance
- Active Learning (AL) reduces labeling costs by selecting a small proportion of samples from unlabeled data for labeling and training
- Deep Active Learning (DAL) maximizes model performance under limited labeling cost/budget
- No available performance evaluation of DAL methods under fair comparison settings
- Authors construct a DAL toolkit called DeepAL+ that re-implements 19 highly-cited DAL methods
- Comparative experiments using frequently used datasets and DAL algorithms conducted
- Factors such as batch size and number of epochs in the training process influence efficacy of DAL
- Paper provides comprehensive survey of deep active learning methods and evaluates their performance under fair comparison settings
- Construction of DeepAL+ toolkit enables researchers to easily access and utilize different DAL methods for their own experiments
- Exploration of influential factors enhances understanding in this area and assists researchers in optimizing their DAL approaches.

## Summary

In their paper titled "A Comparative Survey of Deep Active Learning," authors Xueying Zhan, Qingzhong Wang, Kuan-hao Huang, Haoyi Xiong, Dejing Dou and Antoni B. Chan address the challenge of deep learning's reliance on extensive labeled data for good performance. They highlight how Active Learning (AL) can reduce labeling costs by selecting a small proportion of samples from unlabeled data for labeling and training. To maximize model performance under limited labeling cost/budget, they propose Deep Active Learning (DAL) as a feasible solution. While there have been numerous methods developed for DAL and various literature reviews conducted, the authors note that there is currently no available performance evaluation of DAL methods under fair comparison settings. To fill this gap in research, the authors construct a DAL toolkit called DeepAL+. This toolkit re-implements 19 highly-cited DAL methods. The authors also survey and categorize DAL-related works and conduct comparative experiments using frequently used datasets and DAL algorithms. Additionally, the authors explore factors such as batch size and number of epochs in the training process that influence the efficacy of DAL. By providing insights into these factors, their work aims to offer better references for researchers designing their own DAL experiments or carrying out DAL-related applications. Overall, this paper contributes to the field by providing a comprehensive survey of deep active learning methods and evaluating their performance under fair comparison settings. The construction of the DeepAL+ toolkit enables researchers to easily access and utilize different DAL methods for their own experiments. The exploration of influential factors further enhances understanding in this area and assists researchers in optimizing their DAL approaches.