Iterated distributional and lexicon-driven learning in a symmetric neural network explains the emergence of features and dispersion

We present a neural network model of phonetic and phonological acquisition that can handle two distinct phenomena: category creation and auditory dispersion. Within a single neural network, learning proceeds in two stages. The first stage is distributional learning, during which the model induces phonological features from an auditory input distribution; in the second stage, the model acquires knowledge about the relation between lexical categories and the auditory input distribution. The model can be used bidirectionally: once perceptual learning is complete, the network can also be asked to speak. In the production direction, effortful, perceptually peripheral tokens are avoided. In a chain of iterated learners, in which the output of one generation serves as the input to the next, sound systems emerge that maintain sufficient contrast at a moderate articulatory cost, regardless of the initial distribution.