Implications of learning models for linguistic theory: the richness of memory, the importance of loss, and the emergence of structure

This paper introduces the special issue on implications of learning models, such as deep neural networks and instance-based models, for linguistic theory. This special issue aims to address the core question of linguistic theory – what makes human languages the way they are – by comparing models of language learning (i.e., different theories of the language acquisition device). We bring the papers together by discussing a number of recurrent themes: (i) the importance of rich memory, which enables extensive memorization without impairing generalization; (ii) our increasing understanding of the learning mechanisms that enable the emergence of structure from experience (Hebbian and error-driven); and (iii) the tasks that need to be solved for successful language learning. Limitations and future directions are highlighted.