Input data management in real-time AI systems

Abstract

A real-time AI system in the real world needs to monitor an immense volume of data. To do this, the system must filter out much of the incoming data. However, it must remain responsive to important or unexpected events in the data. This paper describes some simple approaches to data management, shows how they can fail to be both adequately selective and responsive, and presents an approach that improves on the simple approaches by making use of information about the system's resources and ongoing tasks. The new approach has been applied in a system for monitoring patients in a surgical intensive-care unit. 1