Safe Learning-Enabled Systems
As artificial intelligence (AI) systems rapidly increase in size, acquire new capabilities, and are deployed in high-stakes settings, their safety becomes extremely important. Ensuring system safety requires more than improving accuracy, efficiency, and scalability: it requires ensuring that systems are robust to extreme events, and monitoring them for anomalous and unsafe behavior. The objective of the Safe Learning-Enabled Systems program, which is a partnership between the National Science Foundation, Open Philanthropy and Good Ventures, is to foster foundational research that leads to the design and implementation of learning-enabled systems in which safety is ensured with high levels of confidence. While traditional machine learning systems are evaluated pointwise with respect to a fixed test set, such static coverage provides only limited assurance when exposed to unprecedented conditions in high-stakes operating environments. Verifying that learning components of such systems achieve safety guarantees for all possible inputs may be difficult, if not impossible. Instead, a system’s safety guarantees will often need to be established with respect to systematically generated data from realistic (yet appropriately pessimistic) operating environments. Safety also requires resilience to “unknown unknowns”, which necessitates improved methods for monitoring for unexpected environmental hazards or anomalous system behaviors, including during deployment. In some instances, safety may further require new methods for reverse-engineering, inspecting, and interpreting the internal logic of learned models to identify unexpected behavior that could not be found by black-box testing alone, and methods for improving the performance by directly adapting the systems’ internal logic. Whatever the setting, any learning-enabled system’s end-to-end safety guarantees must be specified clearly and precisely. Any system claiming to satisfy a safety specification must provide rigorous evidence, through analysis corroborated empirically and/or with mathematical proof.