Learning in Factored Domains with Information-Constrained Visual Representations