Constrained Machine Learning

Our Adaptive Domain Model approaches carrying a domain’s known structure, a conserved grade, its many dimensions, an equivariance, in its types that directly inform the resulting fitted weights. Our design considers a unique Bayesian approach that allows for inference with confidence intervals, and opens the door for continuous learning.

The payoff compounds: we imagine a typed domain model that is more precise, smaller, and cheap enough for simple hardware, and a constellation of them carries the work a monolithic transformer carries weakly and expensively. It’s a complex landscape, but one we believe is worth exploring theoretically and researching with aim to make it a next-generation intelligence platform.

This section builds that constellation end to end, as a research program and sequence of proposals that fit the theoretical framing. It rests on the ADM pre-print and the framework’s other formal work, and reads the white-box program of Buchanan, Pai, Wang, and Ma and their open CRATE derivation as a specification that in many cases supports our approach, even as we reach beyond their reading.