Current AI systems of both types can excel at many tasks, but they often prove to be brittle when things in the real-world change in ways that were not expected by their designers. In many real-world settings, change and surprises are inevitable, and brittleness can lead to (sometimes catastrophic) failures. In the best case, the change is discovered before anything unplanned or hazardous happens, allowing engineers to avoid costly shutdowns as the AI systems are retrained or reengineered. This inflexibility is particularly problematic for mission-critical applications that require continuous functionality.

Enabling AI systems to adapt rapidly includes three main steps. First, the system will constantly sense the environment and identify when something relevant to its tasks has changed. It then learns about the novelty by characterizing what has changed in the world. Finally, it will adjust its intelligent behavior to the changing world as rapidly as possible. PRIDE with COLTRANE uses probabilistic programming for data-driven machine learning to recognize and characterize the novelty, as well as symbolic AI to flexibly adapt its reasoning. Each type of AI is powerful in its right, but when they are combined, they unlock new capabilities that were previously unachievable.

This foundational technology enables robust, reliable AI that can adapt in the face of change, which is especially helpful when systems cannot be taken out of service for updates. PRIDE with COLTRANE could eventually help incorporate advanced AI into a wider range of crucial activities and situations, such as self-driving ambulances, disaster recovery robots, and defense systems.

Related Contract

Compositionally Organized Learning to Reason about Novel Experiences (COLTRANE)