The Four Big Bets (Illustrated via a Journey in Program Synthesis)

Опубликовано 9 марта 2018, 19:50

Sumit Gulwani will talk about the four bets that he made in the second half of his research career, which yielded greater personal satisfaction. He will illustrate this in the context of his research agenda around Program Synthesis, which is the task of automatically generating a program that meets the user’s specification.
• Customer Connection: Sumit’s mid-life crisis inspired him to spend months studying help forums and interviewing customers to identify the right nuances of problem definitions and its applications. This led to his work on interactive programming-by-examples (PBE), with data wrangling and code refactoring as killer applications.
• Framework: The initial success of Flash Fill (an Excel 2013 feature), which allows users to automate syntactic string transformation tasks from examples, spurred investment into many more task domains. Sumit bet on his theoretical background to invest into the PROSE framework that facilitates algorithmic design and development for various task domains.
• Research and Engineering---better together: To accelerate innovation and its delivery, Sumit committed to engineering product-quality implementation of his research instead of relying on product teams for tech-transfer. He thus found funding to set up a dedicated team of researchers and engineers.
• Cross-disciplinary research (The “AI=PL+ML” equation): Various components in PBE leverage both logical reasoning and creative heuristics. Sumit bet on MSRI collaboration to use ML for efficient development of these heuristics and to make them better and adaptable. Such an integration of PL and ML can have big implications for AI software development in general.
Sumit now thrives on a fresh set of challenges related to data collection (to prioritize investments and improve algorithms) and striking the right balance between long-term and short-term as well as between research and engineering.

See more at microsoft.com/en-us/research/v...