49. WEIGHTED AGGREGATION OF CLASSIFIERS FOR ACTIVE LEARNING

Department: Computer Science & Engineering
Faculty Advisor(s): Yoav Freund

Primary Student
Name: Akshay Balsubramani
Email: abalsubr@ucsd.edu
Phone: 858-729-8393
Grad Year: 2015

Abstract
In modern machine learning applications, labeled data are often expensive or difficult to obtain compared to unlabeled data, which can profoundly limit the accuracy of a conventional learning algorithm. This motivates the paradigm of active learning, in which the algorithm interactively queries the labels of data points which it believes would be most informative, ultimately often requiring far fewer labels for good accuracy. In this research, we devise a novel algorithm, Weighted Active Learning, which essentially uses a weighted majority vote among simple 'expert' classifiers to select points to query. We provide a theoretical analysis of this algorithm and prove that it is significantly more general than existing approaches in several respects. A crowdsourcing application for Twitter posts is also discussed.

Related Links:

  1. http://cseweb.ucsd.edu/~abalsubr/

« Back to Posters or Search Results