Learning to read English requires learning the complex statistical dependencies between orthography and phonology. Previous research has focused on how these statistics are learned in neural network models provided with as much training as needed. Children, however, are expected to acquire this knowledge in a few years of school with only limited instruction. We examined how these mappings can be learned efficiently, defined by tradeoffs between the number of words that are explicitly trained and the number that are correct by generalization. A million models were trained, varying the sizes of randomly selected training sets. For a target corpus of about 3000 words, training sets of 200-300 words were most efficient, producing generalization to as many as 1800 untrained words. Composition of the 300 word training sets also greatly affected generalization. The results suggest directions for designing curricula that promote efficient learning of complex material. [Additional funding by the Vilas Trust at UW-Madison. This paper was published in: "Proceedings of The 41st Annual Meeting of the Cognitive Science Society," edited by C. Freska et al., Cognitive Science Society, 2019.]