Learning Graphical Concepts

Many of the artifacts humans produce exhibit a compositional structure. One example is the graphical model. Machine learning researchers make frequent use of graphical motifs and abstractions such as trees, chains, rings, grids, mixtures, and plates to constrain the space of graphical models they consider. We show how one could learn these motifs and abstractions, or “graphical concepts,” by finding programs that generate common graphical models. In particular, we present the Compositional Exploration/Compression (CEC) algorithm, a general purpose multitask program induction algorithm. THE CEC ALGORITHM