[HOME] [EDUCATION] [PUBLICATIONS] [RESEARCH] [SOFTWARE] [INTERNSHIPS] [PROJECTS] 

LEARNING WITH IMPERFECT/PARTIAL SUPERVISION

Most of the standard off-the-shelf supervised learning algorithms are generally developed for an ideal world. They often make strong assumptions which make them less than ideal for applying them directly to real world problems. For example training points are often noisily labeled,  training samples are not independent and identically distributed, the samples can be biased, is is not clear how to get the objective ground truth, the desired performance metric may be quite different etc. Once I joined the industry and started looking into real data I soon realized that most of the basic assumptions in developing classification algorithms had to be questioned. Suitable modifications had to be made to model these deviations from the ideal scenarios. These resulted in development of new algorithms which gave a significant improvement in performance over off-the-shelf standard classification algorithms. This line of research  has been motivated by problems in medical imaging, specifically in CAD where the task is to build a classifier to predict  whether a suspicious region on a medical image (like a X-ray, CT scan, or MRI) is malignant or benign.

PUBLICATIONS

Learning from crowds Vikas C. Raykar, Shipeng Yu, Linda H. Zhao, Gerardo Valadez, Charles Florin, Luca Bogoni, and Linda Moy, Journal of Machine Learning Research, 11(Apr):1297−1322, 2010 [abstract] [paper] [bib]

Supervised Learning from Multiple Experts: Whom to trust when everyone lies a bit Vikas C. Raykar, Shipeng Yu, Linda Zhao, Anna Jerebko, Charles Florin, Gerardo Valadez, Luca Bogoni, and Linda Moy, In Proceedings of the 26th International Conference on Machine Learning (ICML 2009), pp.889-896, Montreal, June 2009.  [paper] [discussion] [slides] [bib] 

Bayesian Multiple Instance Learning: Automatic Feature Selection and Inductive Transfer Vikas C. Raykar, Balaji Krishnapuram, Jinbo Bi, Murat Dundar, and R. Bharat Rao, In Proceedings of the 25th International Conference on Machine Learning (ICML 2008), pp.808-815, Helsinki, July 2008.  [paper] [slides] [bib]

On Ranking in Survival Analysis: Bounds on the Concordance Index Vikas C. Raykar, Harald Steck, Balaji Krishnapuram, Cary Dehing-Oberije, and  Philippe Lambin. In Advances in Neural Information Processing Systems (NIPS 2007), vol. 20, pp. 1209–1216 , 2008.  [paper] [slides] [spotlight slide] [bib]

The copyrights of publications are with the respective publishers. The papers are being reproduced here for timely dissemination of scholarly information.