Building and using AI engines – what do we do when we’ve machine learned everything?

An AI engine is a system where one collects data from a system, the data from that system are used to improve AI models about the system, those models are redeployed, and evaluated in that system. In this talk I will discuss our efforts to build AI engines for cancer both at the Fred Hutch and at national scale. I will highlight the potential impact of building an AI engine by showing some results from an AI model that can be used to generate gene expression data from experimental design descriptions of those experiments. I will connect these ideas back to one of the grand challenges in modern statistics – how do we do inference when the “data” are generated from a model? I’ll describe our initial efforts toward “inference with predicted data” and highlight growth of this research area within the statistical community. This is joint work with many people at Fred Hutch, Synthesize Bio, Johns Hopkins, Memorial Sloan Kettering, Dana Farber, and at UW statistics.