# Deep Active Learning with Contrastive Learning Under Realistic Data Pool Assumptions

**Authors:** Jihyo Kim, Jeonghyeon Kim, Sangheum Hwang  
**arXiv:** [2303.14433](https://arxiv.org/abs/2303.14433)  
**Date:** 2023-03-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.14433v1/

## Key points

- The paper addresses the complexities of active learning in real-world scenarios with unlabeled data pools that may contain irrelevant or ambiguous samples.
- Traditional active learning methods are often evaluated in ideal settings with only in-distribution samples relevant to the target task.
- The authors propose new active learning benchmarks that include both in-distribution and out-of-distribution samples to address this issue.
- They introduce a novel active learning method that prioritizes acquiring informative in-distribution samples by leveraging labeled and unlabeled data pools and selecting samples based on clusters in the feature space constructed through contrastive learning.
- Experimental results show that the proposed method outperforms existing approaches by requiring a lower annotation budget for the same level of accuracy.
- By considering more realistic assumptions about data distributions, this research contributes to advancing active learning techniques for deep neural networks operating in complex real-world environments.

## Summary

The paper "Deep Active Learning with Contrastive Learning Under Realistic Data Pool Assumptions" by Jihyo Kim, Jeonghyeon Kim, and Sangheum Hwang delves into the complexities of active learning in real-world scenarios where unlabeled data pools may contain irrelevant or ambiguous samples. Traditional active learning methods are often evaluated in ideal settings with only in-distribution samples relevant to the target task. However, in practice, data pools can include out-of-distribution samples that are task-irrelevant or too ambiguous for classification. To address this issue, the authors propose new active learning benchmarks that incorporate both in-distribution and out-of-distribution samples. They introduce a novel active learning method designed to prioritize acquiring informative in-distribution samples. This method leverages both labeled and unlabeled data pools and selects samples based on clusters in the feature space constructed through contrastive learning. Experimental results demonstrate that the proposed method outperforms existing active learning approaches by requiring a lower annotation budget to achieve the same level of accuracy. By considering more realistic assumptions about the diversity of data distributions in unlabeled pools, this research contributes to advancing active learning techniques for deep neural networks operating in complex real-world environments. Overall, this study highlights the importance of adapting active learning strategies to handle diverse and potentially challenging data scenarios, ultimately improving model performance and efficiency in practical applications.