Learning Adaptive Image Representations: Kate Saenko, UMass Lowell

Starts:
1:00 pm on Monday, February 25, 2013
Ends:
2:00 pm on Monday, February 25, 2013
Location:
MCS 148
Abstract: The ideal image representation often depends not just on the task but also on the domain. Recent studies have demonstrated a significant degradation in the performance of state-of-the-art image classifiers when input feature distributions change due to different image sensors and noise conditions, pose changes, a shift from commercial to consumer video, and, more generally, training datasets biased by the way in which they were collected. In this talk, I will describe recent work in learning domain-invariant image representations for multi-class classifiers. First, I will present an algorithm that learns representations which explicitly compensate for domain mismatch. Specifically, it learns a transformation that maps image features from the target (test) domain to the source (training) domain by maximizing similarity between the two domains. I will then discuss several extensions, including ways to discover latent domains and incorporate similarity constraints between points in the target domain, such as similar views or motion tracks. Finally, I will present a modified method for learning image representations that can be efficiently realized as linear classifiers. Learning adaptive representations for linear classifiers is particularly interesting as they are prevalent in vision applications, with fast linear SVMs forming the core of some of the most popular object detection methods. We optimize both the transformation and classifier parameters jointly, and introduce an efficient cost function based on misclassification loss. Our method combines several features previously unavailable in a single algorithm: multi-class adaptation through representation learning, ability to map across heterogeneous feature spaces, and scalability to large datasets. Topic: Object Recognition, Domain Adaptation, Representation Learning Bio: Kate Saenko is an Assistant Professor of Computer Science at the University of Massachusetts Lowell. She received her PhD from MIT, followed by postdoctoral work at UC Berkeley and Harvard. Her research spans the areas of computer vision, machine learning, speech recognition, and human-robot interfaces. She is currently involved in a large multi-institution NSF-sponsored project, conducting research in statistical scene understanding and physics-based visual reasoning. She is also the principal investigator on DARPA’s Mind’e Eye project, developing models for recognizing and describing human activities.