Learning from images by integrating different perspectives
Sally A. Goldman, Sharath R. Cholleti · 2008
While much machine learning research focuses on learning tasks in which one begins from scratch, for many learning scenarios an important component, or even the entire task, is to learn how to combine the predictions made by several independent algorithms or experts. In this dissertation, we study a variety of problems that involve combining predictions from multiple sources. First, we study a type of content-based image retrieval where the user is only interested in a portion of the image. We model this task as a multiple-instance (MI) learning problem and describe a new algorithm, MI-Winnow, that first converts an MI problem into a regular machine learning problem and then uses Winnow to learn. One important component of our work is to combine hypotheses generated using multiple representations. Along with providing experimental evaluation of MI-Winnow, we present some theoretical results for learning in MI setting. In addition, as part of this work, we present a new salient-point representation that reduces the number of salient points by using a segmentation algorithm as a mask but still preserves the overall variety in the image. Next, we study the evaluation of multiple segmentations of the same image using different existing evaluators each being best suited for different types of images. We describe a new meta-learning algorithm to combine the existing base evaluators to create a better evaluator that adaptively weights the base evaluators in response to the specific input image. Finally, we present a new algorithm, Veritas, to estimate the ground truth in medical images from multiple expert segmentations. A very unique aspect of this work is that we perform the combination without any training data, such as labeled images or a ranking of the expert segmentations.