AI Engineering
Realizing the Promise of AI Solutions
Artificial intelligence (AI) holds incredible promise for transforming our software-driven world in general and DoD mission capabilities in particular. AI can compute more data more quickly than even automated software. That computational advantage allows us to make the most of human-machine teaming, freeing humans to focus their attention on the types of tasks they do best.
However, solutions to-date, while often brilliant, are difficult to replicate, verify, and validate. Practiced by virtuosos and talented amateurs, current AI solutions are created using intuition and brute force, with haphazard progress, casual transmission, and extravagant use of available materials.
Leading a movement for AI Engineering
The SEI knows that engineering practice for AI software and data is needed to capitalize on the promise of those capabilities for the nation’s military and government defense and security organizations. Further, the SEI has the experience, skills, and practical creativity to bridge the gap between inventing the possible and deploying the practical.
As a result, the SEI is leading a movement to establish and mature an engineering discipline for the data and software of AI systems. When your organization develops AI-enabled capabilities using AI Engineering practice, you will see
- Rapid, iterative delivery of robust AI-enabled capabilities
- Data management practices tuned for the needs in AI for evidence, traceability, and continuous innovation
- Interface designs for effective human-machine teaming in AI systems
- Continuous acceptance testing
- Transition of tools and critical skills for AI
- Ongoing monitoring and validation to counteract an AI-enabled capability’s enhanced attack surface
- Adaptable, flexible, and evolvable solutions
Featured Work
xView 2 Challenge
The xView 2 Challenge applied computer vision and machine learning to analyze electro-optical satellite imagery before and after natural disasters to assess building damage. The competition’s sponsor was the Department of Defense’s Defense Innovation Unit (DIU). This technology is being used to assess building damage from wildfires in Australia and the United States.
Designing Trustworthy Artificial Intelligence
The Human-Machine Teaming Framework guides development in creating Artificial Intelligence systems that are accountable to humans, cognitive of speculative risks and benefits, secure, and usable.
Verifying and Validating Mission-Critical Systems Using Quantum Computing
The SEI is working to study how quantum computing can serve as the next paradigm for optimization in fields like software verification and validation.
Explainable AI: Why Did the Robot Do That?
To help human users trust their robot team members in critical situations, we develop tools that allow autonomous systems to explain their behavior.
Featured Publications
Threats to Machine Learning Applications
August 18, 2020 • Presentation
Mark Sherman
This presentation illustrates where machine learning applications can be attacked, the means for carrying out the attacks, and some mitigations that can be employed.
read
AI Engineering: 11 Foundational Practices
September 12, 2019 • White Paper
This initial set of recommendations can help organizations that are beginning to build, acquire, and integrate artificial intelligence capabilities into business and mission systems.
read
Machine Learning in Cybersecurity: A Guide
September 05, 2019 • Technical Report
Jonathan SpringJoshua FallonApril Galyardt
This report suggests seven key questions that managers and decision makers should ask about machine learning tools to effectively use those tools to solve cybersecurity problems.
read
Deepfakes—What Can Really Be Done Today?
August 27, 2019 • Video
Rotem D. GuttmanZachary Kurtz
Rotem Guttman and Zach Kurtz explain what deepfakes are, how they work, and what kind of content it's possible to create with current techniques and technology.
watch
Learning by Observing via Inverse Reinforcement Learning
March 22, 2019 • Video
Ritwik GuptaEric Heim
This SEI Cyber Talk episode explains how inverse reinforcement learning can be effective for teaching agents to perform complex tasks with many states and actions.
watch