Jon Rawski

Affiliation: 
Stony Brook University
Talk Title: 
Homeostatic Reinforcement Learning for Generative Grammars
Event Type: 
Brown Bag Talk
Semester: 
Fall 2016
Date: 
Wednesday, September 28, 2016, 1:00 pm

Abstract:

Reinforcement Learning (RL), despite being one of the most widely used and neurologically robust machine learning algorithms, has an uneasy history with linguistics. Specifically, the requirement of an internal, restricted hypothesis space and other learnability restraints is inadequately satisfied by externally defined "naive" reward. However, recent insights from computational neuroscience offer a possible bridge. In this talk, I will discuss the concept of biological homeostasis, and how it fits into RL. I will show that linguistic knowledge can be conceived as a homeostatic grammar space, where deviations from an attractor point yield corrective responses. To demonstrate that homeostasis is a true linguistic Third Factor, and not a consequence of a particular language model, I reproduce a P&P syntax acquisition result of Yang (2002), and also simulate late-L2 resistance to constraint demotion in Harmonic Grammar. I will conclude with some tentative hypotheses of homeostasis in bilinguals. Apart from interesting models and simulations, this approach offers prospects for uniting ideas from neural and linguistic theory in order to provide a more coherent explanatory neurolinguistics.