Instructors:  Hal Daumé III, Naomi Feldman & Ellen Lau
 Mondays 2:00pm - 4:30pm
Location:  MMH 1108B
Course #s:
 LING 848 & CMSC 818

About this class: Much recent work in cognitive neuroscience demonstrates that brains are engaged in predicting future input, such as upcoming words in a speech or text stream. Thirty years of research in language modeling attempts to do the same thing computationally, using models that capture various aspects of language: local configurations (n-gram models), syntactic constraints or semantic preferences. These are implemented variously as lookup-tables, parsers, recurrent neural networks, etc. This course is about understanding both sides of the prediction task: both how the brain does it, what aspects are most salient to the brain, and how this connects with computational models designed to capture various linguistic structures.

Prerequisites: This course should be accessible to students who have taken: CMSC/LING723, LING646, LING689/889, or any rigorous course in probability/statistics or machine learning. Please talk to one of us if you are in doubt.

Course structure, assignments and grading: (Due to Lillian Lee and modified) For each lecture, there are two readings, and every student must closely read one of the two papers and write a research proposal based on the reading (roughly 2-3 paragraphs; include idea/hypothesis, dataset or human experiment); you must also skim the other paper. You must post your research proposal to Piazza, where you will get feedback from us and from classmates before class. During class, we discuss/defend/refine the proposals in class. So your job is to practice generating lots of research ideas and getting lots of feedback, and there's lots to talk about, with common ground to the discussion.  A sample of one of these proposals can be found on the "course administration" page of one of Lillian's courses. At the end of the semester, you will work in groups to put together a more extensive research proposal on a topic of your choosing. Each group must include at least one person coming in with more 'brain'/cognitive expertise and at least one coming in with more 'computational' expertise.