Artificial Intelligence (AI) and Machine Learning (ML) investment is exponentially increasing in both the DoD and industry, the DoD requested $130.1 billion in 2023 alone. However, much of this investment goes towards early-stage research and development or proof of concept projects which fail to reach warfighters. Often work in early-stage R&D succumbs to the valley of death due gaps in funding for operationalization of lab proven concepts. One such hurdle often encountered when attempting to operationalize AI/ML models is the lack of end-to-end infrastructure supporting all AI/ML life cycle stages such as development, testing, deployment, and maintenance.

End to end infrastructure is important for AI/ML models due to their unique data needs, however, this necessity is rarely mentioned in literature. Without robust end to end architecture, adeptly tracking and monitoring model parameters and data is often a cumbersome labor-intensive process. This lack of tracking and slow triaging of model performance issues often results in subpar model performance when fielded in operational settings. This paper looks at the critical components required to bridge the R&D valley of death to operationalize AI/ML based on a currently fielded rapid retraining architecture for object recognition. The paper starts by defining and describing three main stages in an end-to-end rapid retraining pipeline: 1) data collection and preprocessing; 2) experimentation and model development; and 3) model deployment and monitoring. From here the paper will cover a use case using the pipeline, showing greater than 25% reductions in model retraining time and increases in model accuracy using a mix of commercial off the shelf (COTS) and custom pipeline tools. Ultimately, work in this paper demonstrates the importance of automating the AI/ML development and deployment process to effectively operationalize this critical new tool for the warfighter.

Keywords
AI, ISR SYSTEMS

Additional Keywords