Analyzing the Stability of Convolutional Neural Networks against Image Degradation

Hamed Habibi Aghdam, Elnaz Jahani Heravi, Domenec Puig

2016

Abstract

Understanding the underlying process of Convolutional Neural Networks (ConvNets) is usually done through visualization techniques. However, these techniques do not provide accurate information about the stability of ConvNets. In this paper, our aim is to analyze the stability of ConvNets through different techniques. First, we propose a new method for finding the minimum noisy image which is located in the minimum distance from the decision boundary but it is misclassified by its ConvNet. Second, we exploratorly and quanitatively analyze the stability of the ConvNets trained on the CIFAR10, the MNIST and the GTSRB datasets. We observe that the ConvNets might make mistakes by adding a Gaussian noise with s = 1 (barely perceivable by human eyes) to the clean image. This suggests that the inter-class margin of the feature space obtained from a ConvNet is slim. Our second founding is that augmenting the clean dataset with many noisy images does not increase the inter-class margin. Consequently, a ConvNet trained on a dataset augmented with noisy images might incorrectly classify the images degraded with a low magnitude noise. The third founding reveals that even though an ensemble improves the stability, its performance is considerably reduced by a noisy dataset.

Paper Citation

in Harvard Style

Habibi Aghdam H., Jahani Heravi E. and Puig D. (2016). Analyzing the Stability of Convolutional Neural Networks against Image Degradation . In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 4: VISAPP, (VISIGRAPP 2016) ISBN 978-989-758-175-5, pages 370-382. DOI: 10.5220/0005720703700382

in Bibtex Style

@conference{visapp16,
author={Hamed Habibi Aghdam and Elnaz Jahani Heravi and Domenec Puig},
title={Analyzing the Stability of Convolutional Neural Networks against Image Degradation},
booktitle={Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 4: VISAPP, (VISIGRAPP 2016)},
year={2016},
pages={370-382},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0005720703700382},
isbn={978-989-758-175-5},
}

in EndNote Style

TY - CONF
JO - Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 4: VISAPP, (VISIGRAPP 2016)
TI - Analyzing the Stability of Convolutional Neural Networks against Image Degradation
SN - 978-989-758-175-5
AU - Habibi Aghdam H.
AU - Jahani Heravi E.
AU - Puig D.
PY - 2016
SP - 370
EP - 382
DO - 10.5220/0005720703700382