PAC Confidence Sets for Deep Neural Networks via Calibrated Prediction

Sangdon Park; Osbert Bastani; Nikolai Matni; Insup Lee

PAC Confidence Sets for Deep Neural Networks via Calibrated Prediction

Machine Learning 2020-02-18 v2 Machine Learning

Authors: Sangdon Park , Osbert Bastani , Nikolai Matni , Insup Lee

Abstract

We propose an algorithm combining calibrated prediction and generalization bounds from learning theory to construct confidence sets for deep neural networks with PAC guarantees---i.e., the confidence set for a given input contains the true label with high probability. We demonstrate how our approach can be used to construct PAC confidence sets on ResNet for ImageNet, a visual object tracking model, and a dynamics model for the half-cheetah reinforcement learning problem.

Keywords

pac confidence sets deep neural networks calibrated prediction generalization bounds resnet imagenet

Cite

@article{arxiv.2001.00106,
  title  = {PAC Confidence Sets for Deep Neural Networks via Calibrated Prediction},
  author = {Sangdon Park and Osbert Bastani and Nikolai Matni and Insup Lee},
  journal= {arXiv preprint arXiv:2001.00106},
  year   = {2020}
}

Comments

Accepted to ICLR 2020

Related papers

View all related →