Training accurate visual classifiers from large data sets critically depend on learning the right representation for the problem. I will discuss a representation learning framework based on an iterative interaction of two components: a feature generator suggesting candidate features, and a feature selector choosing among them. In the feature selector role, I will present a feature selection algorithm for Support Vector Machines (SVMs) enabling selection among hundreds of thousands of features, while keeping an accuracy comparable to computationally expensive wrapper methods. For the feature generator, I will discuss two main examples: part-based features generation for human detection, and sparse feature generation for object recognition under severe test-time speed constraints. In both examples state of the art classifiers (at the time of submission) were learned. Specifically the sparse classifiers are currently the state of the art for visual classification with tight computational budget.
Aharon Bar Hillel is a researcher in Microsoft Research ATLI (Advanced Technical Labs Israel). He received his PH.D thesis from The Hebrew university of Jerusalem in 2006, dealing with machine learning and computer vision topics. Since then He has been doing machine learning and computer vision oriented research in Intel Research (2006-2008) and in GM Research (2009-2012). He is interested in learning representation for machine learning tasks, including distance function learning, feature selection and synthesis, and deep learning.