During their talk, they shared how Domino and NVIDIA are helping them create an MLOps framework to bring development, security, and operations best practices to machine learning and achieve a 10 to 100X efficiency gain for data scientists.
Here are three examples they shared:
Reducing supply chain risks. Lockheed Martin is in the process of building the Orion crew module for human spaceflight that is targeted to go to the moon in 2024. A project of this type requires many different unique components and there's significant risk if a part can not be supplied on time. To reduce risk of delays, data scientists are building models with Domino and NVIDIA that analyze open source unstructured data, such as information about a merger between two suppliers; classify and assess potential risks; and provide insight on what might happen next (for example, if a company may experience financial losses). These insights are combined with extensive internal knowledge of the company’s supply chain to give supply chain professionals a better sense of any risks so they can respond to them.
With Domino and NVIDIA, the company has been able to train these models with effectively no training data—something that required a large amount of compute power--and rapidly deploy interfaces so they could create an active learning loop to obtain feedback from individuals on how the models are performing.
“The ability to kind of keep track of that data and assign it to these projects really made this project a success.”
—Mike Johnson, Lead Data Scientist, Lockheed Martin
Building trust in AI systems through its participation in the AlphaDogFight competition, hosted by DARPA. Participants were tasked to develop an AI fighter system that could compete with humans in a Top Gun style aircraft fight. Lockheed was one of many teams, and it secured second place by taking a reinforcement learning approach and using Domino and NVIDIA to spin up multiple training jobs (each of which took on the order of 30 days) and track experiments.
“Before we had Domino, it took eight weeks to get access to a GPU. Now it's just a push button away. Those savings add up across thousands of employees at scale, and that really makes a difference.”
—Greg Forrest, Senior Manager of AI and AiMLabs, Lockheed Martin
Predicting equipment failures before they affect operations. In his final example, Mike Johnson shared what it takes to build an F35 aircraft and the number of different operations that have to happen (all in one mile-long building) to turn raw materials into an aircraft. To minimize equipment downtime, the company is prototyping an AI-driven system that can continuously monitor every machine for potential issues and can notify staff when machines aren’t operating properly. So far, using Domino and NVIDIA, around three machine learning engineers were able to develop and train over 250 models and deploy 150 models at the same time.
About the Speakers
Mike Johnson is the technical lead for a team of data scientists, machine learning engineers, and data engineers to deliver AI solutions across Lockheed Martin. He has built machine learning solutions in numerous fields including manufacturing optimization, semiconductor reliability, human resources, radar signal analysis, and time series search. While he has an undying love of natural language processing, he has recently been focused on applying deep learning at scale to the domain of unsupervised anomaly detection.
Greg Forrest is responsible for leading the development of Lockheed Martin’s corporate artificial intelligence strategy, creating an enterprise AI ecosystem and MLOps pipeline, and leading teams that develop transformative AI and machine learning capabilities for Lockheed Martin and its customers.