Latent Task Adaptation with Large-Scale Hierarchies

TitleLatent Task Adaptation with Large-Scale Hierarchies
Publication TypeConference Paper
Year of Publication2013
AuthorsJia, Y., & Darrell T.
Other Numbers3617

Recent years have witnessed the success of large-scaleimage classification systems that are able to identify objects among thousands of possible labels. However, it isyet unclear how general classifiers such as ones trained onImageNet can be optimally adapted to specific tasks, eachof which only covers a semantically related subset of allthe objects in the world. It is inefficient and suboptimalto retrain classifiers whenever a new task is given, and isinapplicable when tasks are not given explicitly, but implicitly specified as a set of image queries. In this paper wepropose a novel probabilistic model that jointly identifiesthe underlying task and performs prediction with a linear-time probabilistic inference algorithm, given a set of queryimages from a latent task. We present efficient ways to estimate parameters for the model, and an open-source toolboxto train classifiers distributedly at a large scale. Empiricalresults based on the ImageNet data showed significant performance increase over several baseline algorithms


This work was partially supported by funding provided through National Science Foundation grants IIS : 1212928 ("Reconstructive recognition: Uniting statistical scene understanding and physics-based visual reasoning") and IIS-1116411 ("Hierarchical Probabilistic Layers for Visual Recognition of Complex Objects"). Additional support was provided by DARPA's Minds Eye and MSEE programs and the Toyota Corporation. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors or originators and do not necessarily reflect the views of the funders.

Bibliographic Notes

Proceedings of the International Conference on Computer Vision 2013 (ICCV 2013), Sydney, Australia

Abbreviated Authors

Y. Jia and T. Darrell

ICSI Research Group


ICSI Publication Type

Article in conference proceedings