Computer Science Department
School of Computer Science, Carnegie Mellon University
Interactive Machine Learning in Diamond
Unindexed search systems, such as Diamond, are more useful than indexed search systems precisely when the cost of indexing cannot be amortized and classifiers are inexpensive to create. This thesis establishes the latter condition for many image classification tasks. To accommodate a wide variety of visual phenomena, a flexible, learned image representation, Semantic Texton Forests, is adapted for use in Diamond. To reduce the amount of interaction required to produce a high-quality classifier, a novel active learning algorithm, Active Learning by Measure Approximation, is theoretically developed. To consolidate all components of the system, a usable interface, Algum, is implemented. The result is an effective workflow realizing the Diamond vision of iterated, interactive hypothesis exploration.