Naturalness and Bimodality of Software, and the practical use of Language Models
When: September 06, 2023, 14:00 - 15:30
Where: ECHO-ARENA and collegerama
CS Distinguished Speaker Lectures (CS-DSL)
Abstract
After discovering, back in 2011, that Language Models are useful for modeling repetitive patterns in source code (c.f. The “Naturalness” of software https://dl.acm.org/doi/10.5555/2337223.2337322), and exploring some applications thereof, more recently (since about 2019) our group at UC Davis has focused on the observation that Software, as usually written, is bimodal, admitting both the well-known formal, deterministic semantics (mostly for machines) and probabilistic, noisy semantics (for humans). This bimodality property affords both new approaches to software tool construction (using machine-learning) and new ways of studying human code reading. In this talk, I’ll give an overview of the Naturalness/Bimodality program, some very recent work on “bimodal prompting”, and finally some recent directions on evaluating the quality of code produced by large language models.
Speaker Bio
Prem Devanbu received his B.Tech from IIT Madras, and a Ph.D from Rutgers University. After working in Industrial software development at Bell Laboratories and offshots in New Jersey, he joined UC Davis where is now a Distinguished Research Professor. His papers have won several awards, including multiple best paper awards, distinguished paper awards, most influential paper awards, and test-of-time awards. Three of his papers were invited to appear in CACM Research Highlights. He served as PC Chair of ESEC/FSE 2006 and ICSE 2010, and also as GC of MSR 2014 and ESEC/FSE 2020. He has served on the Editorial boards of ACM TOSEM, IEEE ToSE, the JSME, and the EMSE Journal; he serves currently on the CACM Editorial Board. In 2021, he was awarded the ACM SIGSOFT Outstanding Research Award, and in 2022 he was awarded the Alexander von Humboldt Research Award. He is an ACM Fellow.
Recording
