Faculty Recruiting Support CICS

Machine Learning for Databases: We're Doing it Wrong

20 Nov
Wednesday, 11/20/2019 11:00am to 12:00pm
Computer Science Building, room 151
Speaker:  Ryan Marcus

The database research community has recently seen a groundswell of new research integrating machine learning techniques into database components (e.g., query optimizers, index selection, cardinality estimation). In this talk, I will take a step back and first examine the fundamental reasons why machine learning techniques can improve data management systems, and how those reasons can inform good choices when designing ML solutions for systems problems. Through the concepts of inductive bias and biased parameter maps, I will argue that many recent works (including my own) fall drastically short of achieving their full potential. Finally, I will show how we put these concepts into practice through our (far from perfect) learned query optimizer, Neo.

Ryan Marcus is a postdoc researcher at MIT, working under Tim Kraska on learned components for database systems. Ryan recently graduated from Brandeis University, where he studied applications of machine learning to cloud databases under Olga Papaemmanouil. Before that, Ryan took courses in gender studies and mathematics at the University of Arizona, while banging his head against supercomputers at Los Alamos National Laboratory. He enjoys long walks down steep gradients, and shorter walks down gentler ones. He's also looking for a job! You can find Ryan online at https://rmarcus.info

Faculty Host