StreamlinedML

ABSTRACT

Lockheed Martin Advanced Technology Laboratories developed a prototype Machine Learning (ML) ecosystem for the AFRL StreamlinedML program. This ecosystem will significantly improve the end-to-end process of configuring, training, evaluating, and deploying machine learning models, and enable ML DevOps for AF/DoD applications.

WHY?

ML model development for AF/DoD applications is often characterized by a narrow set of experts, with domain knowledge and ML experience, exploring solutions to pressing operational problems in isolation. The model needs an integration ecosystem that democratizes ML, which will allow wider adoption of ML that lowers the barriers to entry for use.

WHAT?

We are building a composable ML ecosystem that includes:

1. ML model and technology encapsulation using microservices and containerization
2. A runtime framework that enables extensibility through orchestration, including model-as-a-service (MaaS) capabilities
3. ML workflows to capture ML abstractions and provide a basis for the runtime composition
4. An integrated ML repository that allows storage, query, and manipulation of algorithms/models, metadata, (hyper) parameters, and provenance
5. Shared access to foster a scalable ML (SML) community

HOW?

We will develop and deploy a prototype SML ecosystem that provides SDKs, toolkits, and services to access and develop best-in-class algorithms by using a wide variety of ML stacks and datasets. A prototype version is currently deployed in the AF HPC and AWS environments; SML users have started evaluating the APIs and framework.