Modular Meta-cognition (ModMetaCog)
Abstract
The Modular Meta-cognition (ModMetaCog) program aims to develop machine introspection technologies that enable AI systems to analyze and explain their behavior and improve their performance. Current AI systems sometimes work well in the lab but then perform poorly in the field where they operate as part of larger systems that behave unpredictably and/or when they encounter situations not anticipated during training. ModMetaCog will develop and apply recent results in AI/ML saliency and explainability to create introspection capabilities for AI systems, in effect wrapping a second AI system around the original system. The second AI system acts to analyze and explain the behavior of the first, identify anomalies and potential problems, and either acts to make changes to the first system to fix the problem or to communicate its findings to a human operator for action. ModMetaCog will develop both basic theory of meta-cognition and prototype introspection modules for specific types of AI problems, such as classification, detection, control, and others. ModMetaCog will create machine introspection technologies that enable AI/ML systems to identify when they are performing poorly and to improve their performance when required.
Document Details
- Document Type
- Accomplishment
- Publication Date
- Oct 01, 2024
- Source ID
- 022058dbd00a72a929a130b56f0e6fd6