One common loss function in neural network classification tasks is Categorical Cross Entropy (CCE), which punishes all misclassifications equally. However, classes often have an inherent structure. For instance, classifying an image of a rose as ``violet'' is better than as ``truck''. We introduce SimLoss, a drop-in replacement for CCE that incorporates class similarities along with two techniques to construct such matrices from task-specific knowledge. We test SimLoss on Age Estimation and Image Classification and find that it brings significant improvements over CCE on several metrics. SimLoss therefore allows for explicit modeling of background knowledge by simply exchanging the loss function, while keeping the neural network architecture the same. Code and additional resources are available at https://github.com/konstantinkobs/SimLoss
Description
SimLoss: Class Similarities in Cross Entropy | SpringerLink
%0 Conference Paper
%1 kobs2020simloss
%A Kobs, Konstantin
%A Steininger, Michael
%A Zehe, Albin
%A Lautenschlager, Florian
%A Hotho, Andreas
%B Foundations of Intelligent Systems
%C Cham
%D 2020
%E Helic, Denis
%E Leitner, Gerhard
%E Stettinger, Martin
%E Felfernig, Alexander
%E Raś, Zbigniew W.
%I Springer International Publishing
%K 2020 cross entropy function loss myown
%P 431--439
%T SimLoss: Class Similarities in Cross Entropy
%U http://arxiv.org/abs/2003.03182
%X One common loss function in neural network classification tasks is Categorical Cross Entropy (CCE), which punishes all misclassifications equally. However, classes often have an inherent structure. For instance, classifying an image of a rose as ``violet'' is better than as ``truck''. We introduce SimLoss, a drop-in replacement for CCE that incorporates class similarities along with two techniques to construct such matrices from task-specific knowledge. We test SimLoss on Age Estimation and Image Classification and find that it brings significant improvements over CCE on several metrics. SimLoss therefore allows for explicit modeling of background knowledge by simply exchanging the loss function, while keeping the neural network architecture the same. Code and additional resources are available at https://github.com/konstantinkobs/SimLoss
%@ 978-3-030-59491-6
@inproceedings{kobs2020simloss,
abstract = {One common loss function in neural network classification tasks is Categorical Cross Entropy (CCE), which punishes all misclassifications equally. However, classes often have an inherent structure. For instance, classifying an image of a rose as ``violet'' is better than as ``truck''. We introduce SimLoss, a drop-in replacement for CCE that incorporates class similarities along with two techniques to construct such matrices from task-specific knowledge. We test SimLoss on Age Estimation and Image Classification and find that it brings significant improvements over CCE on several metrics. SimLoss therefore allows for explicit modeling of background knowledge by simply exchanging the loss function, while keeping the neural network architecture the same. Code and additional resources are available at https://github.com/konstantinkobs/SimLoss},
added-at = {2021-01-24T18:36:41.000+0100},
address = {Cham},
author = {Kobs, Konstantin and Steininger, Michael and Zehe, Albin and Lautenschlager, Florian and Hotho, Andreas},
biburl = {https://www.bibsonomy.org/bibtex/2eb6e8ed0d018b2baae1df7f216cdcacb/hotho},
booktitle = {Foundations of Intelligent Systems},
description = {SimLoss: Class Similarities in Cross Entropy | SpringerLink},
editor = {Helic, Denis and Leitner, Gerhard and Stettinger, Martin and Felfernig, Alexander and Ra{\'{s}}, Zbigniew W.},
interhash = {c9693bde2183871e64852bd39bcf396b},
intrahash = {eb6e8ed0d018b2baae1df7f216cdcacb},
isbn = {978-3-030-59491-6},
keywords = {2020 cross entropy function loss myown},
pages = {431--439},
publisher = {Springer International Publishing},
timestamp = {2021-01-24T22:43:11.000+0100},
title = {SimLoss: Class Similarities in Cross Entropy},
url = {http://arxiv.org/abs/2003.03182},
year = 2020
}