Testing using Privileged Information (TUPI)

by Adapting Features with Statistical Dependence

ICCV 2021





How might we improve a predictor after it has been trained and without having access to the original training data or labels?
We propose a new algorithm that can improve given predictros using privileged information available at test time.


Abstract

Given an imperfect predictor, we exploit additional features at test time to improve the predictions made, without retraining and without knowledge of the prediction function. This scenario arises if training labels or data are proprietary, restricted, or no longer available, or if training itself is prohibitively expensive. We assume that the additional features are useful if they exhibit strong statistical dependence to the underlying perfect predictor. Then, we empirically estimate and strengthen the statistical dependence between the initial noisy predictor and the additional features via manifold denoising. As an example, we show that this approach leads to improvement in real-world visual attribute ranking.




Testing using Privileged Information (TUPI) scenarios are more common than you might think!






Bibtex Reference

@inproceedings{KT2021:ICCV,
author = {Kwang In Kim and James Tompkin},
title = {Testing using Privileged Information by Adapting Features with Statistical Dependence},
booktitle = {Proc. ICCV},
pages = {9405--9413},
year = {2021},
}

Related Projects


Acknowledgements

This work was supported by the National Research Foundation of Korea (NRF) grant (No. 2021R1A2C2012195), Institute of Information and communications Technology Planning and evaluation (IITP) grant (2021--0--00537, Visual Common Sense Through Self-supervised Learning for Restoration of Invisible Parts in Images), and IITP grant (2020--0--01336, Artificial Intelligence Graduate School Program, UNIST), funded by the Korea government (MSIT). This material is based on research sponsored by Defense Advanced Research Projects Agency (DARPA) and Air Force Research Laboratory (AFRL) under agreement number FA8750-19-2-1006. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of Defense Advanced Research Projects Agency (DARPA) and Air Force Research Laboratory (AFRL) or the U.S. Government.