Continuous Manifold Based Adaptation for Evolving Visual Domains
Title | Continuous Manifold Based Adaptation for Evolving Visual Domains |
Publication Type | Conference Paper |
Year of Publication | 2014 |
Authors | Hoffman, J., Darrell T., & Saenko K. |
Other Numbers | 3692 |
Abstract | We pose the following question: what happens when testdata not only differs from training data, but differs from itin a continually evolving way? The classic domain adaptation paradigm considers the world to be separated intostationary domains with clear boundaries between them.However, in many real-world applications, examples cannot be naturally separated into discrete domains, but arisefrom a continuously evolving underlying process. Examples include video with gradually changing lighting andspam email with evolving spammer tactics. We formulate anovel problem of adapting to such continuous domains, andpresent a solution based on smoothly varying embeddings.Recent work has shown the utility of considering discretevisual domains as fixed points embedded in a manifold oflower-dimensional subspaces. Adaptation can be achievedvia transforms or kernels learned between such stationarysource and target subspaces. We propose a method to consider non-stationary domains, which we refer to as Continuous Manifold Adaptation (CMA). We treat each targetsample as potentially being drawn from a different subspaceon the domain manifold, and present a novel technique forcontinuous transform-based adaptation. Our approach canlearn to distinguish categories using training data collectedat some point in the past, and continue to update its modelof the categories for some time into the future, without receiving any additional labels. Experiments on two visualdatasets demonstrate the value of our approach for severalpopular feature representations. |
Acknowledgment | This research was supported in partby DARPA Minds Eye and MSEE programs, by NSFawards IIS-0905647, IIS-1134072, and IIS-1212798, andby support from Toyota. |
Bibliographic Notes | Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Columbus, Ohio, June 2014 |
Abbreviated Authors | J. Hoffman, T. Darrell, and K. Saenko |
ICSI Research Group | Vision |
ICSI Publication Type | Article in conference proceedings |