Abstract: 

In this talk, Elliot Branson, Director of Machine Learning & Engineering at Scale, will explain how efficient forms (without SGD, for example) of weak learning can effectively enable unsupervised learning. He’ll compare different approaches to weak labeling and feature engineering with Snorkel, FlyingSquid and open source packages like scikit-learn and XGBoost. He’ll then demonstrate how human-in-the-loop QA can deliver highly accurate models. When used together with weak learning, this approach offers performance which is the best-of-both worlds. The resulting ML models perform better than weak learning or human labelling alone.

Bio: 

Elliot Branson Is the Director of AI and Engineering at Scale and leads the Machine Learning, AI Infrastructure, Platform, Federal, 3D, and Mapping products. In his prior work, he helped create the Cruise Automation self-driving car and served as the first Head of Perception and AI. His interest in robotics and AI started with national and international robotics competitions in high school and continued in college and grad school where he published on field robotics, localization, computer vision, and AI systems. His previous work includes stints on the Google Project Tango AR platform and Air Force MURI research programs.