The Weekly Newsletter of MIT Linguistics

Ling-Lunch 11/03 — Jon Rawski

Speaker: Jon Rawski (Stony Brook)
Tittle: Homeostatic Reinforcement Learning for Harmonic Grammars
Date/Time: Thursday, November 3/12:30pm-1:50pm
Location: 32-D461

The main idea of this talk is to bridge a particularly thorny divide between linguistics and neuroscience. Reinforcement Learning (RL), despite being one of the most widely used and neurologically robust learning algorithms, has an uneasy history with generative grammar. Specifically, the requirement of an internal, restricted hypothesis space and other learnability restraints is inadequately satisfied by externally defined “naive” reward (Chomsky 1959).

Reparation of RL and linguistics is made urgent by the discovery that: 1) phonology is at most a regular language (Kaplan & Kay 1994, Heinz 2011), meaning it is restricted to finite-state automata, and 2) RL is perfectly computed by cortical neurons (Schultz et al 1997). One recent attempt is Charles Yang’s (2002) “Naïve Parameter Learner”, which uses RL to successfully model acquisition of overt [WH-movement] and [V2] parameters, yet fails to provide more than an ad-hoc definition for “reward”.

In this talk I show that recent insights from computational neuroscience offer a possible strategy. A recent framework called Homeostatic Reinforcement Learning (HRL) (Keramati and Gutkin 2014) treats “reward” as an internal satisfaction of multiple, parallel constraints in a homeostatic space. This immediately suggests Harmonic Grammar. I posit that the weighted constraints in Harmonic Grammar constitute a homeostatic space, and the Harmony function is a necessary and sufficient condition for RL in constraint-based grammars. I then show that this model successfully learns final obstruent devoicing in Russian, among others. I conclude with some tentative hypotheses for homeostasis in bilinguals and in late-L2 learners. 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.