Braden Hancock
Stanford (CS PhD Candidate)

About the talk:

Training accurate classifiers requires many labels, but each label provides only limited information (one bit for binary classification). In this work, we propose BabbleLabble, a framework for training classifiers in which an annotator provides a natural language explanation for each labeling decision. A semantic parser converts these explanations into programmatic labeling functions that generate noisy labels for an arbitrary amount of unlabeled data, which is used to train a classifier. On three relation extraction tasks, we find that users are able to train classifiers with comparable F1 scores from 5-100 times faster by providing explanations instead of just labels. Furthermore, given the inherent imperfection of labeling functions, we find that a simple rule-based semantic parser suffices.

The full paper can be found here: https://arxiv.org/abs/1805.03818.

Video:

To access the live webcast of the talk (active at 16:28 of the day of the presentaton) and the archived version of the talk, use the URL SU-EE380-20190220. This is a first class reference and can be transmitted by email, Twitter, etc.

A URL referencing a YouTube view of the lecture will be posted here a week or so following the presentation.

About the Speaker

Braden is in his final year of a computer science PhD at Stanford. His research has focused on building new supervision interfaces and interaction modes with machine learning systems, including various flavors of supervising with natural language. His goal is to make it possible for non-ML experts to create near state-of-the-art machine learning systems for new tasks in hours instead of months. He has worked at various times at Facebook Research, Google Research, MIT Lincoln Laboratory, Johns Hopkins University, and the Air Force Research Laboratory. For more information on current and past projects, see his personal website: www.bradenhancock.com.