Hasty: A Generative Model Compiler

DARPAs Data Driven Discovery of Models (D3M) program aimed to automate machine learning so as to allow non-expert government users to pose queries, make predictions, and otherwise model data without necessarily having a background in data science or machine learning. Our research was to develop an extensive set of model primitives and contribute them to a library of discoverable component models that would be assembled by other teams pipeline search and tuning software tools. Our part of this work was developing a software toolchain that made it possible to vastly extend a set of such model primitives. Our contributions included significantly influencing the design of the D3M primitive interface, implementing the original core primitives that demonstrated this interface, implementing an additional number of specialized primitives, and explaining the opportunity, untapped, for program participants to construct their own primitives based on our developed and contributed generative model compiler technology.