Discriminatively trained neural classifiers can be trusted, only when the
input data comes from the training distribution (in-distribution). Therefore,
detecting out-of-distribution (OOD) samples is very important to avoid
classification errors. In the context of OOD detection for image
classification, one of the recent approaches proposes training a classifier
called "confident-classifier" by minimizing the standard cross-entropy loss on
in-distribution samples and minimizing the KL divergence between the predictive
distribution of OOD samples in the low-density regions of in-distribution and
the uniform distribution (maximizing the entropy of the outputs). Thus, the
samples could be detected as OOD if they have low confidence or high entropy.
In this paper, we analyze this setting both theoretically and experimentally.
We conclude that the resulting confident-classifier still yields arbitrarily
high confidence for OOD samples far away from the in-distribution. We instead
suggest training a classifier by adding an explicit "reject" class for OOD
samples.
Description
[1904.12220] Analysis of Confident-Classifiers for Out-of-distribution Detection
%0 Journal Article
%1 vernekar2019analysis
%A Vernekar, Sachin
%A Gaurav, Ashish
%A Denouden, Taylor
%A Phan, Buu
%A Abdelzad, Vahdat
%A Salay, Rick
%A Czarnecki, Krzysztof
%D 2019
%K uncertainty
%T Analysis of Confident-Classifiers for Out-of-distribution Detection
%U http://arxiv.org/abs/1904.12220
%X Discriminatively trained neural classifiers can be trusted, only when the
input data comes from the training distribution (in-distribution). Therefore,
detecting out-of-distribution (OOD) samples is very important to avoid
classification errors. In the context of OOD detection for image
classification, one of the recent approaches proposes training a classifier
called "confident-classifier" by minimizing the standard cross-entropy loss on
in-distribution samples and minimizing the KL divergence between the predictive
distribution of OOD samples in the low-density regions of in-distribution and
the uniform distribution (maximizing the entropy of the outputs). Thus, the
samples could be detected as OOD if they have low confidence or high entropy.
In this paper, we analyze this setting both theoretically and experimentally.
We conclude that the resulting confident-classifier still yields arbitrarily
high confidence for OOD samples far away from the in-distribution. We instead
suggest training a classifier by adding an explicit "reject" class for OOD
samples.
@article{vernekar2019analysis,
abstract = {Discriminatively trained neural classifiers can be trusted, only when the
input data comes from the training distribution (in-distribution). Therefore,
detecting out-of-distribution (OOD) samples is very important to avoid
classification errors. In the context of OOD detection for image
classification, one of the recent approaches proposes training a classifier
called "confident-classifier" by minimizing the standard cross-entropy loss on
in-distribution samples and minimizing the KL divergence between the predictive
distribution of OOD samples in the low-density regions of in-distribution and
the uniform distribution (maximizing the entropy of the outputs). Thus, the
samples could be detected as OOD if they have low confidence or high entropy.
In this paper, we analyze this setting both theoretically and experimentally.
We conclude that the resulting confident-classifier still yields arbitrarily
high confidence for OOD samples far away from the in-distribution. We instead
suggest training a classifier by adding an explicit "reject" class for OOD
samples.},
added-at = {2019-12-07T00:49:56.000+0100},
author = {Vernekar, Sachin and Gaurav, Ashish and Denouden, Taylor and Phan, Buu and Abdelzad, Vahdat and Salay, Rick and Czarnecki, Krzysztof},
biburl = {https://www.bibsonomy.org/bibtex/2d0df65aea33eed66752a52b362d04a0c/kirk86},
description = {[1904.12220] Analysis of Confident-Classifiers for Out-of-distribution Detection},
interhash = {862989308b5a908947afba2aad42ac92},
intrahash = {d0df65aea33eed66752a52b362d04a0c},
keywords = {uncertainty},
note = {cite arxiv:1904.12220Comment: SafeML 2019 ICLR workshop paper},
timestamp = {2019-12-07T00:49:56.000+0100},
title = {Analysis of Confident-Classifiers for Out-of-distribution Detection},
url = {http://arxiv.org/abs/1904.12220},
year = 2019
}