About

Talk

Optimizing Training for Sparse Workloads in Tensorflow

Level: Beginner+

Twitter is a company with massive amounts of data. Thus, it is no wonder that the company applies machine learning in myriad of ways. In this session, we will be exploring some of the challenges Twitter faced by working with heavily text-based (sparse) data and some of the improvements we have made in our Tensorflow-based platform to deal with these use cases. This session will focus on explaining, in depth, the use case of one of teams that benefits from our Machine Learning platform: Timelines Ranking. We plan to discuss its feature pipeline, modeling decisions as well as platform improvements. In the modeling side, we will be discussing hyperparameter tuning as well as different architecture explorations (alongside discretization and isotonic calibration).  In the platform side, we will be exploring some of the challenges Twitter faced by working with heavily text-based (sparse) data and some of the improvements we have made in our Tensorflow-based platform to deal with these use cases. Overall, we plan to give a holistic view into one of Twitter’s most prominent use cases.