Efficient Causal Discovery and Abstraction in Perception Streams

Human discovery of cause and effect in perception streams requires reliable online inference in highly unstructured noisy environments with very few (maybe even only one) positive examples. Automated causal discovery research, on the other hand, has typically operated in a qualitatively different setting: Data used for learning is cast into more static scenarios where all prospect observable causes are known in advance. Typically when learning it is assumed that all data is known (i.e., learning is done in batch mode), and little work has been done on the important but hard topic of causal abstraction. In this paper, we present a system that is a capable of wading through an unknown set of possible causes (events) in an online fashion to identify a small set of plausible candidates for some effect of interest. As new events are encountered, they can be added to the system in an online fashion. We provide a mechanism to learn arbitrary noisy logic formulae and present a method to do causal abstraction.