Humans-in-the-Loop Don't Perform Gradient Descent
Speaker

Akshat is currently an engineering manager at Scale AI, where he leads a team that is building software and systems to label text, audio and computer vision ground truth data for ML teams with incredible accuracy. He previously has worked at D. E. Shaw and database and fintech startups, and studied computer science and mathematics at MIT. A former competitive programmer, he was the first competitor from India to receive a gold medal at the IOI.
SUMMARY
Data labeling at scale is a complicated endeavor involving operations, machine learning and game theory. In this talk we focus on the game theory, and introduce a model of human-in-the-loop labeling as an optimization problem. Based on our (Scale AI’s) experience of operating tens of thousands of labelers on high complexity tasks, we will discuss how various problems predicted by the model map to observed results.

