Artificial Intelligence Reliability and Traceability (AIRT)
Abstract
The Artificial Intelligence Reliability and Traceability (AIRT) program will develop design-time and run-time technologies to ensure the correct functioning of AI-enabled systems. As AI deployment scales up, it becomes more important for machine learning (ML) systems and their training data to be explainable, which means providing rationale for classifications, characterizing confidence level of the classifications, and, as a consequence, conveying an understanding of how the system will behave with similar inputs. Explainability, however, is not sufficient to ensure that ML systems meet reliability requirements and are free of bias, in the sense that the ML operates consistently with domain-focused predictive models, nor that they are traceable, in the sense that there are mappings between the models and the ML behaviors. AIRT will develop the test, evaluation, verification, and validation (TEVV) technologies that system developers need to ensure that AI-enabled systems will correctly perform their intended functions. The AIRT TEVV technologies will address the challenge of how to specify AI-related behaviors and then how to verify the specified behaviors using both analytic formal approaches, which emphasize mathematical modeling and reasoning, and traditional statistical-sampling based approaches. AIRT will also develop design principles for machine learning and related systems that enhance reliability and traceability without appreciable compromise to reasoning capability. Additionally, AIRT will develop traceability approaches that model the learning behavior of an AI component to enable developers, testers, and operators to gain detailed knowledge of how the AI system reached a computational state. The AIRT program aims to make the design and operation of AI systems more scientific and safe.
Document Details
- Document Type
- Accomplishment
- Publication Date
- Oct 01, 2023
- Source ID
- 214d867c28675d3ecbfbd88d3dc4b9ff