Track: Emerging Analytics
Engineering Implementation of AI Principles
Wednesday, April 14, 2-2:40pm EDT
Much of the day-to-day use of AI in the past decade has been in the consumer applications on the internet (e.g. simple tasks such as question-answering, product recommendations, etc.) where the consequence of an AI output is relatively minor. With increasing desire to adopt AI in business and mission critical applications, there are major concerns about reliability, trust and ethics underlying these AI applications.
To address these concerns, many government and private organizations (e.g. High-Level Expert Group on AI of the European Commission, Defense Innovation Board of the US Department of Defense, etc.) have defined AI principles to provide guidance for the creation, operation and sustenance of AI systems. These cover wide range of topics such as, appropriate human oversight at all levels, transparency & accountability, privacy & data governance, robustness & safety, etc. As of now, these AI principles are just expressions of desired goals, with no engineering practices to support demonstration of conformance.
The purpose of this talk is to explore the various implications of the AI Principles and develop a measurement framework to help with the realization of trustworthy AI systems. Since Machine Learning is a statistical technique and relies heavily on data, the measurements of AI systems have to reflect that property, similar to Statistical Process Control, commonly used in manufacturing. Metrics have to reflect the variance of the outputs for given inputs, as well as the variance of the outputs for different input data sets in a normalized manner to be meaningful. Such a framework has to include more rigorous engineering practices and analysis of engineering artifacts to demonstrate the fit-for-purpose of AI systems.