The Center for Cognitive Computation (CCC) invites you to the online lecture of Yanli Zhou (New York University). Zoom link will be sent around on the department's Talks mailing list. If missed, please ask Mihály Bányai.
Abstract: Humans are efficient learners of functions - the ability to represent and compose functions emerges early in development. Work in linguistics further suggests that humans are capable of learning much more complex function interactions. Going back to Kiparsky (1968), linguists have cataloged numerous linguistic phenomena covering four
logical patterns for ordering two interactive functions. “Feeding” in a context-free grammar is when one function creates the context for another to operate, and “counterfeeding” is the converse. “Bleeding” occurs when the operation of one function removes the context for another, and “counterbleeding” is its converse. In this project, our aim is to determine the extent to which humans and computational models can learn to compose functions based on the system of interactions. We introduce a learning task that adapts and extends the Piantadosi & Aslin (2016) design, evaluating participants on zero-shot function compositions covering all four interaction types. Our findings indicate that participants can correctly infer the underlying functions based on limited input-output examples; they can also
generalize to novel combinations of functions across different conditions. Close examinations of participants’ response patterns reveal a number of potential inductive biases. Furthermore, we demonstrate that a standard sequence-to-sequence transformer model can be trained to complete the same task with high levels of accuracy via meta-learning. Incorporating guidance from human data, our model can learn to reproduce behavioral patterns that mirror the complete and complex way humans perform function compositions.