Human adults have rich theories in their heads of how the world works. These theories include objects and relations for both concrete and abstract concepts. Everything we know either must be innate or learned through experience. Yet it's unclear how much of this model needs to be innate for a computer. The core question this dissertation addresses is how a computer can develop rich relational theories using only its raw sensor data. We address this by outlining a "bridge" between raw sensors and a rich relational theory. We have implemented parts of this bridge, with other parts as feasibility studies, while others remain conceptual.

At the core of this bridge is Ontol, a system that constructs a conceptual structure or "ontology" from feature-set data. Ontol is inspired by cortical models that have been shown to be able to express invariant concepts, such as images independent of any translation or rotation. As a demonstration of the utility of the ontologies created by Ontol, we present a novel semi-supervised learning algorithm that learns from only a handful of positive examples. Like humans, this algorithm doesn't require negative examples. Instead, this algorithm uses the ontologies created by Ontol from unlabeled data, and searches for a Bayes-optimal theory given this "background knowledge".

The rest of the dissertation shows in principle how Ontol can be used as the "workhorse" for a system that finds analogies, discovers useful mappings, and might ultimately create theories, such as a "gisty" theory of "number".