Vidit Jain
Yahoo! Labs, Bangalore
Date: Wednesday, April 3rd, 2013
Time: 4 PM
Venue: CS102.
Supervised learning relies on the assumption of similarity between the distribution of training and test instances. However in practice there are often significant differences between these distributions. These differences arise due to the cost of collecting large training data sets and also due to the difficulties in obtaining training instances from a particular target test domain. How can we rapidly and simply adapt the trained models to a new test distribution, even when we do not have access to the original training data? In this talk, we will explore techniques based on Gaussian process regression to facilitate such adaptation. In particular, we will present our observations in two application areas. First we will discuss the problem of re-ranking image search results, where we achieve a significant gain in relevance over a commercial search engine. Then we will present our solution for domain adaptation for face detection. These two results have been published at the WWW and CVPR conferences, respectively.
Vidit Jain is a scientist in Yahoo! Labs Bangalore. He received his MS and PhD degrees from University of Massachusetts Amherst and a B.Tech degree from IIT Kanpur, all in computer science. He has worked at several research institutes in the US, Europe, and India including the research labs at Yahoo!, Microsoft, and Kodak. His research has involved building statistical machine learning models for different tasks including ranking of search results, face analysis, and semantic organization of noisy, short text-segments (e.g., see Yahoo! Amoeba). He has contributed to several research publications, technical manuscripts, and patent filings in these areas.