Vision (TIN2006-15308-C02)

Multitask Learning is a novel machine learning approach that learns each problem better by also learning from the training signals of other related problems. Whereas learning at the base-level focuses on accumulating experience on a specific learning task, learning at this meta-level is concerned with accumulating experience on the performance of a learning system applied to different related problems. This project addresses the development of multitask learning methods for a Computer Vision learning scenario: the classification of visual data in images. In order to prove the potential and the generalization ability of the multitask paradigm, we will work on two different computer vision fields: object recognition and medical image analysis. Our main hypothesis is that multitask learning will provide two complementary advantages to practical solutions of the studied problems: more classification accuracy when classifiers are trained on several problems in parallel (parallel knowledge transfer), and less training requirements when classifiers are trained sequentially (sequential knowledge transfer). In the first case, we expect large accuracy increments, and in the second case significant reductions in the number of training samples that are necessary for learning at a given performance level. The application part of the project splits in 4 subtasks: a) an automatic system for facial detection and identification, b) a system for automatic traffic sign recognition, c) a computer-aided system for plaque characterization in ultrasound images, and d) an intelligent system for automatic annotation of intestine motility in wireless endoscopy