Event Location: NSH 1507
Bio: Alex Berg’s research concerns computational visual recognition. He has worked on general object recognition in images, action recognition in video, human pose identification in images, image parsing, face recognition, image search, and machine learning for computer vision. He is currently an assistant professor in computer science at Stony Brook University. Prior to that he was a research scientist at Columbia University and Yahoo! Research. His PhD at U.C. Berkeley developed a novel approach to deformable template matching. He earned a BA and MA in Mathematics from Johns Hopkins University and learned to race sailboats at SSA in Annapolis.

Abstract: Machine learning has produced amazingly efficient algorithms for training classifiers with enormous amounts of data. Computer vision researchers need to determine what data representation, and more importantly, what labels to feed these ravenous learning engines. I will discuss these questions in the context of our recent work on efficient additive classifiers that make possible experiments on
classifying images into more than 10,000 categories, the largest scale experiments of this type. These in turn reveal structure in image classification that is not present in previous, smaller, experiments. In particular we find evidence of a correlation between a version of semantic similarity and the difficulty of visual
classification.

Complementing that work using categories as labels, we also consider using visual attributes as labels. Attribute labels are more flexible than category labels, and can be composed to provide information about multiple levels of categories in a hierarchy or even about individuals. Our work on face verification demonstrates
that using attributes as an intermediate representation can not only generalize to unseen classes, but can produce state of the art results on a widely attacked problem. Toward increasing the usefulness of attributes for recognition, we present a technique for automatically identifying visual attributes by mining noisy web data.

Joint work with: Tamara Berg, Subhransu Maji, Neeraj Kumar, Peter
Belhemeur, Shree Nayar, Jonathan Shih, Jia Deng, & Fei-Fei Li