Scholarlink

Future intelligence, surveillance, reconnaissance, and targeting capabilities will rely on Artificial Intelligence/Machine Learning (AI/ML) to support time critical missions including indications and warning intelligence preparation of the battlespace, and real-time targeting. Historically, exploitation and analysis of imagery have relied entirely on the expertise of trained analysts. Today, however, the volume of sensor data and the shortened timelines for decisions are driving us towards greater automation. At the same time advances in AI/ML capabilities are providing tools to assist analysts in meeting these missions. For many basic image analysis tasks, such as object detection or object classification, multiple models might be available to process imagery for a given mission. Which model is the best choice for the mission? Our research demonstrates that seemingly similar deep learning methods can yield different performance results. Furthermore, our investigations demonstrate differences due to the choice of training data. The team propose methods for evaluating the training data and the model framework relative to the mission imagery, which leads to a natural strategy for model selection. Using several popular object detection methods applied to multiple data sets, we will present:

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Training the Right ML Model and Training the ML Model Right

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • John M. Irvine,
  • Nazario Irizarry,
  • Franck Olivier Ndjakou Njeunje,
  • Samuel Vilt

摘要

Future intelligence, surveillance, reconnaissance, and targeting capabilities will rely on Artificial Intelligence/Machine Learning (AI/ML) to support time critical missions including indications and warning intelligence preparation of the battlespace, and real-time targeting. Historically, exploitation and analysis of imagery have relied entirely on the expertise of trained analysts. Today, however, the volume of sensor data and the shortened timelines for decisions are driving us towards greater automation. At the same time advances in AI/ML capabilities are providing tools to assist analysts in meeting these missions. For many basic image analysis tasks, such as object detection or object classification, multiple models might be available to process imagery for a given mission. Which model is the best choice for the mission? Our research demonstrates that seemingly similar deep learning methods can yield different performance results. Furthermore, our investigations demonstrate differences due to the choice of training data. The team propose methods for evaluating the training data and the model framework relative to the mission imagery, which leads to a natural strategy for model selection. Using several popular object detection methods applied to multiple data sets, we will present: