"Controlled Rearing of Language Models can Reveal Linguistic Insight" - Kanishka Misra

Originally presented on: Friday, October 11th, 2024 at 12:30pm CT, TTIC, 6045 S. Kenwood Avenue, 5th Floor, Room 530
Title: "Controlled Rearing of Language Models can Reveal Linguistic Insight"
Speaker: Kanishka Misra, TTIC
Abstract: Neural network based language models (LMs) have proliferated across a multitude of disciplines. Meanwhile, their precise role in advancing our understanding of human linguistic knowledge remains a topic of much debate. What can these black boxes tell us about language learning, use, and generalization?
In this talk I will discuss an approach my colleagues and I have taken to contribute to this broad debate. In particular, I will introduce the paradigm of “controlled rearing” where we perform systematic interventions on a learner’s exposure environment, informed by specific hypotheses, and measure the consequent effect on its behavior. I will then describe a case study where we applied the controlled rearing method to understand and analyze the conditions that enable generalization to a rare - and sometimes completely novel - linguistic phenomenon. Specifically, I will show that LMs can learn from indirect evidence, where they demonstrate generalization to a rare/novel phenomenon when their input consists of linguistically related utterances, in a manner that is not explained simply by shallow processing. I will conclude by describing multiple ways in which controlled rearing could be a general toolkit to understand the role of input in facilitating generalization.
#lms #languagemodels #artificialintelligence #machinelearning #computervision #research