Testing AI Learning in Open-world Novelty Scenarios (TALONS)

Navatek proposes to develop scientifically-based verification and validation (V&V) methods for artificial intelligence systems in novel situations. In our approach, V&V of novelty begins with a characterization of novelty along eight distinct dimensions. With this characterization, it is possible to systematically create novel scenarios, assign variables and controls, and test hypotheses as one would in a design of experiments grounded in the scientific method. Navatek proposes to assess the feasibility of our new V&V methods on a self-driving vehicle system. One of the most challenging impediments to widespread adoption of consequential AI such as self-driving vehicles will be convincing stake-holders, including the general public, that these systems can be trusted when human life is at stake. If successful, the technology created in our proposed project could be used within the larger regulatory framework to certify whether an autonomous car, taxi, or truck passes safety standards, similar to how vehicles are graded for crash safety and how compliance with emissions standards are certified. The same technology can be used to certify new military AI systems.