# Deep Active Learning for Scientific Computing in the Wild

**Authors:** Simiao Ren, Yang Deng, Willie J. Padilla, Leslie Collins, Jordan Malof  
**arXiv:** [2302.00098](https://arxiv.org/abs/2302.00098)  
**Date:** 2023-01-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.00098v1/

## Key points

- Deep learning (DL) has revolutionized how researchers approach complex problems in scientific computing.
- Active learning is a promising solution to address the data gap issue, allowing for more efficient data acquisition and model training.
- Scientific computing tasks are often dominated by regression problems, which differ from image classification tasks typically studied in deep active learning (DAL) literature.
- The impact of unknown hyperparameters on DAL performance is a crucial aspect often overlooked in existing literature.
- Some DAL models may underperform when crucial hyperparameters such as pool ratio (γ) are not known beforehand.
- Diversity in sampled data within x-space is highlighted as a key factor contributing to robustness in DAL for scientific computing problems.
- Certain DAL approaches consistently outperformed random sampling even in uncertain conditions, challenging conventional wisdom surrounding uncertainty-based DAL models.
- Developing a comprehensive benchmark dataset for DAL in scientific computing and providing open access to datasets and code sets a new standard for evaluating and improving DAL methods.

## Summary

In the rapidly evolving field of scientific computing, deep learning (DL) has emerged as a game-changer, revolutionizing how researchers approach complex problems. One key challenge in this domain is the data gap that often arises due to the high cost associated with simulations or experiments. To address this issue, active learning has been identified as a promising solution, allowing for more efficient data acquisition and model training. However, while active learning has shown promise in various domains, the deep active learning (DAL) literature has primarily focused on image classification tasks and pool-based methods that may not directly translate to scientific computing problems. In particular, scientific computing tasks are often dominated by regression problems where there is no predefined 'pool' of unlabeled data to draw from. In a groundbreaking study, researchers have delved into the robustness of DAL methods specifically tailored for scientific computing challenges. By evaluating ten state-of-the-art DAL methods across eight benchmark datasets, they have shed light on an important aspect often overlooked in existing literature: the impact of unknown hyperparameters on DAL performance. One crucial finding from their analysis is that some DAL models may underperform when crucial hyperparameters such as pool ratio (γ) are not known beforehand. Surprisingly, many DAL methods exhibited no significant improvement over random sampling in scenarios where ideal pool sizes were uncertain. However, amidst these findings, certain DAL approaches consistently outperformed random sampling even in uncertain conditions. Moreover,the study highlights the importance of diversity in sampled data within x-space as a key factor contributing to robustness in DAL for scientific computing problems. This insight challenges conventional wisdom surrounding uncertainty-based DAL models and underscores the significance of leveraging diverse information sources to enhance model performance. By developing a comprehensive benchmark dataset for DAL in scientific computing and providing open access to both datasets and code for reproducibility purposes,this research sets a new standard for evaluating and improving DAL methods in real-world applications. Ultimately, these findings pave the way for more effective utilization of deep active learning techniques in addressing complex scientific challenges.