Structure and abstraction in phonetic computation: Learning to generalise among concurrent acquisition problems

Sound systems vary dramatically in their lower-level details as a result of cultural evolution, but the presence of systematic organisation is universal. Why does variation pattern differently at these two levels of abstraction, and what can this tell us about the cognitive mechanisms that underpin human acquisition of speech? We explore an evolutionary rationale for the proposal that human learning extends to, and is perhaps even specialised for, making inferences at the higher-order level of abstraction. The ability to infer systematicity from distributional cues, by identifying signatures of structural homogeneity and anticipating subtle exceptions, can bootstrap lower-level learning, and is not subject to the moving target problem, a major evolutionary objection to specialisation in speech cognition. We examine this idea from a statistical perspective, by studying the representational assumptions that underpin generalisation among concurrent phonetic category induction problems. We present a probabilistic model for jointly inferring individual sound classes and a system-wide blueprint for the balance of shared and idiosyncratic structure among these classes. These models lead us to an evolutionary conjecture: culture pushes cognitive adaptation up the hierarchy of abstraction in learning