Events
CDS Seminar: Flexible Inference in Noisy-channel Comprehension
Speaker: Thomas Clark (NYU)
Location: 60 Fifth Avenue, Room 7th Floor Open Space
Date: Friday, September 25, 2026
Abstract: What explains the remarkable robustness of human language to errors, constraints, and adverse conditions during communication, or humans' ability to infer meanings from highly out-of-distribution input like a language they've never studied? In this work, I first present a model of human noisy-channel language comprehension as sequential, approximate, Monte Carlo inference and demonstrate that it captures known human behavioral patterns. Next, I extend this model to the case of cross-language intercomprehension, showing that iterative probabilistic inference and symbolic rules help align the predictions of an untrained, monolingual language model with human behavior in zero-shot translation tasks between unfamiliar but related languages. Taken together, the results provide interpretable explanations for the cognitive strategies that underlie successful communication under real-world constraints.
Bio: I'm Thomas Hikaru Clark, a Faculty Fellow in the Center for Data Science at New York University. I completed my PhD at the Massachusetts Institute of Technology in the Department of Brain and Cognitive Sciences. My research interests include modeling human communication under atypical conditions and constraints, and how human language processing interacts with broader cognitive problem-solving abilities, across settings such as the reading of anomalous sentences, mutual intelligibility between languages, and language production under limited vocabularies. Before my PhD, I was a high school math and computer science teacher, and maintain a strong interest in pedagogy and science education.