Lalit Jain will give a talk entitled “Large Scale Human Driven Data Collection for Preference Learning.”
Modern machine learning crucially relies on features of items to make inferences and predictions. Often times these features are derived from very expensive feature engineering processes, or through unsupervised deep learning methodologies. However, it’s crucial that the features are rich enough to make inferences about human preferences. In this talk, we describe a general methodology of learning representations of items in Euclidean space from similarity judgments using ordinal embedding techniques and extensions to metric learning. Though ordinal embedding has been a popular representational learning technique for over 40 years, there has been little to no analysis of the number of samples needed to learn a high-fidelity model. We provide the first practical results on the number of samples, discuss practical implementations, and also manage to show that ordinal embedding data is enough to completely give a geometric characterization of a set of items. Finally, I will discuss the potential of adaptive data collection techniques.
Lalit Jain is currently a member of Professor Kevin Jamieson’s group. Before coming to UW, he was a postdoc at the University of Michigan mentored by Professor Anna Gilbert. His PhD in Mathematics was advised by Professor Jordan Ellenberg and Professor Robert Nowak.
Previously, he was a Teach for America corps member in San Francisco at Ida B. Wells Continuation High School. He used to teach for the San Francisco and Oakland Math Circles and he was an organizer for Madison Math Circles. Educational equity is something he is very passionate about and loves to discuss.
Learn more about the Community Seminar series, as well as future lectures, here.
This event is open to the public.