Artificial intelligence systems employed for safety and mission-critical decision support are not useful if results arrive after the time of need, regardless of the quality of outputs. Computable artificial intelligence is AI designed with deterministic, efficient processes and algorithms that enable the exact calculation of response times from query to predictions for real-time operations.
Imagine, a fighter pilot has just been scrambled to engage in air combat. Several threats have been identified by ground control and air tasking orders have been sent from command and control (C2) determined by their current information (static tasking). Artificial intelligence has made it into the cockpit, enabling enhanced situational awareness, predictive analytics, automated threat detection, assistance with weapon systems, and increased flight safety, which allows the pilot to make better decisions and react more quickly to threats.
As the battlespace changes and the pilot loses connectivity to the C2 center, they look to the AI onboard for decision support on the best course of action. The AI must analyze a constant stream of incoming data from available sensors to assess the situation, identify the enemy aircraft and its characteristics, predict its future path and behavior, and provide these predictions to the pilot in real time. A split-second decision is required to both keep the pilot safe and ensure mission success. Not only is there no room for error, but there is also no room for latency.
Computability may not be as big a factor in using AI for well-defined problems, but it is crucial in enabling artificial intelligence for real-time operations. Many dynamic, high-stakes AI applications, like decision-support for air combat, require real-time or near-real-time predictions as well as top-tier accuracies. This is difficult to achieve as one often comes at the cost of the other. This problem presents a roadblock to incorporating artificial intelligence/machine learning across many of these challenging use cases.
Popular approaches for complex problem-solving, such as neural networks, often require a large number of calculations to produce predictions. The time it takes to produce a prediction can vary depending on the complexity of the input data and the size of the network as well as hardware requirements, which can be computationally expensive and make information difficult to process at the edge. This can be a significant issue in applications that require real-time predictions or where the response time is critical.
Real-time, computable AI can be used across industries, such as Defense, Healthcare, Finance, etc. to provide quick and accurate insights and responses in high-pressure and time-sensitive situations, improving decision-making and transforming operations. The ability to calculate exact response times combined with the ability to adapt to new data sets in the field on a low SWaP-c system allows for highly dynamic and advanced decision-making for next-gen systems and exponentially increases the value of implementing AI for high-stakes predictions.
Truly computable AI requires determinism that eliminates randomness in both the function and execution of algorithms. While computability seemingly presents a significant roadblock to the widespread use of AI, advancements in computational power and an industry focus on real-time, lifelong learning hold promise for the future.
References:
1 DARPA - Real Time Machine Learning (RTML)
2 ZDNET - Machine learning is going real-time: Here's why and how
