Discovering basic kernels in the first layer of a CNN

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F61988987%3A17610%2F18%3AA1901U95" target="_blank" >RIV/61988987:17610/18:A1901U95 - isvavai.cz</a>

Výsledek na webu

—

DOI - Digital Object Identifier

—

Jazyk výsledku

angličtina

Název v původním jazyce

Discovering basic kernels in the first layer of a CNN

Popis výsledku v původním jazyce

In our research, we are focused on a semantic description of feature maps and corresponding to them kernels that are used in Convolutional neural networks (CNN). A CNN is a hierarchically ordered computational tool that uses a training process to learn and extract abstract characteristics (features) of input objects. The extracted features have to approximate the original object sufficiently for a next stage (classification, regression, etc.). The first layer of a CNN extracts features by performing convolution operations with a number of kernels. Therefore, to study the semantic description of feature maps in the first layer, we should be focused on convolutional kernels (otherwise, weight vectors). We make a conjecture that in the first layer, a CNN learns the most typical image processing convolutional kernels, and they appear in subsequent layers as well. To confirm our claim, we have used 5 CNNs~cite{vgg,inception,resnet,mobilenet,alexnet} trained on the textit{ImageNet} dataset. In the first layer of selected CNNs, we have identified (after performing clustering) the following kernels: gradient kernels with various rotations, Gaussian kernels and texture extracting kernels (e.g., Gabor filters). Further, we have found kernels sensitive to certain color(s) (combinations) that might be dominant in the ImageNet.To conclude, our hypothesis seems to be correct, and a CNN indeed learns standard image processing convolution kernels. A relationship between kernels in further layers is a subject of the future research.

Název v anglickém jazyce

Discovering basic kernels in the first layer of a CNN

Popis výsledku anglicky

In our research, we are focused on a semantic description of feature maps and corresponding to them kernels that are used in Convolutional neural networks (CNN). A CNN is a hierarchically ordered computational tool that uses a training process to learn and extract abstract characteristics (features) of input objects. The extracted features have to approximate the original object sufficiently for a next stage (classification, regression, etc.). The first layer of a CNN extracts features by performing convolution operations with a number of kernels. Therefore, to study the semantic description of feature maps in the first layer, we should be focused on convolutional kernels (otherwise, weight vectors). We make a conjecture that in the first layer, a CNN learns the most typical image processing convolutional kernels, and they appear in subsequent layers as well. To confirm our claim, we have used 5 CNNs~cite{vgg,inception,resnet,mobilenet,alexnet} trained on the textit{ImageNet} dataset. In the first layer of selected CNNs, we have identified (after performing clustering) the following kernels: gradient kernels with various rotations, Gaussian kernels and texture extracting kernels (e.g., Gabor filters). Further, we have found kernels sensitive to certain color(s) (combinations) that might be dominant in the ImageNet.To conclude, our hypothesis seems to be correct, and a CNN indeed learns standard image processing convolution kernels. A relationship between kernels in further layers is a subject of the future research.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10102 - Applied mathematics

Projekt

—

Návaznosti

S - Specificky vyzkum na vysokych skolach

Rok uplatnění

2018

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

Proceedings of the 19th International Student Conference on Applied Mathematics and Informatics

ISBN

9788074641121

ISSN

—

e-ISSN

—

Počet stran výsledku

1

Strana od-do

51-51

Název nakladatele

University of Ostrava

Místo vydání

Ostrava

Místo konání akce

Malenovice

Datum konání akce

10. 5. 2018

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—