EVENT CANCELLED: Neural network models for language acquisition

Event cancelled: A fascinating open seminar by guest speaker Dr Micha Elsner on neural network models for language acquisition.

Dr Micha Elsner is an Associate Professor at the Department of Linguistics at The Ohio State University. He has recently been awarded a Google Research Award for his work on cognitively inspired deep Bayesian neural networks for unsupervised speech recognition. Dr Elsner builds computational models of infant language acquisition, especially  phonetics, phonology and morphology.

Phonetics and phonology are among the first things that infants learn about their language, suggesting that they apply rapid and powerful mechanisms to discover abstract representations from language exposure. Computational models of this process can test our hypotheses about how they do this.

But we don't yet know the answer--- no computer program is yet capable of learning phonetics simply by listening to raw audio. One challenge is the complex, highly correlated nature of audio signals. Another is separating phonologically relevant differences in speech from differences in speaker, speech style and other sources of variation. I will discuss my lab's ongoing effort to understand this process, and our attempt to model it with neural networks, a popular framework for computational learning models. Our models use cognitively plausible goals for their learning, such as memorising what they hear or predicting what they will hear next, to develop abstract representations of speech. By pursuing these goals, they acquire some linguistically meaningful features from audio data, but there is a long way still to go; I will discuss both our successes and challenges so far. 

Although the talk will focus on computer models of learning, Dr Micha Elsner hopes to abstract away from the technical details and explain our high-level ideas about how language and learning work. The talk will present joint work with Cory Shain (PhD student), Stephanie Antetomaso (PhD Student), Naomi Feldman (Associate Professor at the University of Maryland), and Kasia Hitczenko (Graduate student).