Radiotekhnika
Publishing house Radiotekhnika

"Publishing house Radiotekhnika":
scientific and technical literature.
Books and journals of publishing houses: IPRZHR, RS-PRESS, SCIENCE-PRESS


Тел.: +7 (495) 625-9241

 

A convolutional fuzzy neural network for classification tasks

Keywords:

K.P. Korshunova – Post-graduate Student, Department of Computer Facilities, The Branch of National Research University “Moscow Power Engineering Institute” in Smolensk E-mail: ksenya-kor@mail.ru


Convolutional Neural Networks (CNN) are powerful approaches to solve image classification problems. The CNN architectures make the explicit assumption that the inputs are images, which allows to encode certain abstract properties into the architecture. To enable a system to deal with cognitive uncertainties in a manner more like humans, one may incorporate the concept of fuzzy logic into the neural networks. The resulting hybrid system is called fuzzy neural, neural fuzzy, neuro-fuzzy or fuzzy-neuro network. A model of Convolutional Fuzzy Neural Network (CFNN) for classification of real world objects and scenes in images is proposed in the paper. The proposed model\'s architecture is built up of four main types of layers: Convolutional Layer, Pooling Layer, Self-Organization (or Fuzzy) Layer, and Fully-Connected Layer. To form a full Convolutional Fuzzy Neural Network architecture we stack these three parts: 1) a convolutional network (Convolutional and Pooling Layers); 2) The Self-Organization Layer (The Fuzzy Layer); 3) a classifier (some Fully-Connected Layers). The convolutional network (1) takes an input images and form some abstract high-level properties of it by the series of convolutional and pooling layers interchange. The outputs of the Fuzzy Layer (2) neurons represents the values of the mem-bership functions for the fuzzy clusters of input data. These values goes to the input of a classifier (3). Its output is the full CFNN output (the class scores). The Training of The Convolutional Fuzzy Neural Network consists of three independent steps for three components of The Net. First of all we train the convolutional network to form some abstract properties of the input image by backpropagation. Nowadays there is a lot of «pretrained» models that have been already trained on a large data set from a related domain. So we can skip this step of training the full CFNN. The second part is the tuning of the fuzzy layer parameters. The Fuzzy Layer is self-organizing. It is trained in an unsupervised way using a competitive learning scheme. The classifier is trained by a standart backpropagation algorythm. So, the proposed model combines the power of convolutional neural networks and fuzzy logic and is capable of handling uncertainty and impreciseness. The experimental work to measure the effectiveness of CFNN is performing now.
References:

 

  1. Fuler R. Neural fuzzy systems. Abo: Publishing House Abo Akademi University. 1995.
  2. Borisov V.V., Kruglov V.V., Fedulov A.S. Nechetkie modeli i seti. M.: Gorjachaja linija – Telekom. 2007. 287 s.
  3. Keller J.M., Hunt D.J. Incorporating fuzzy membership function into the perceptron algorithm // IEEE Transactions on Pattern Analysis and Machine Intelligence. 1985. V. 7. № 6. P. 693–699.
  4. Mitra S., Pal S. K. Fuzzy multi-layer perceptron, inferencing and rule generation // IEEE Transactions on Neural Networks. 1995. V. 6. P. 51–63.
  5. Mitra S., Pal S. K. Fuzzy self organization, inferencing and rule generation // IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans. V. 26. P. 608–620.
  6. Pal S. K., Mitra S. Multilayer perceptron, fuzzy sets and classification // IEEE Transactions on Neural Networks. 1992. V. 3. № 5. P. 683–697.
  7. Le Cun Y., Boser B., Denker J.S., Henderson D., Howard R.E., Hubbard w., Jackel L.D. Handwritten Digit Recognition with a Back-Propagation Network // Advances in Neural Information Processing Systems 2 (NIPS). 1989. P. 396–404.
  8. Le Cun Y., Bottou L., Bengio Y., Haffner P. Gradient-Based Learning Applied to Document Recognition // Proceedings of the IEEE. 1998. V. 86(11). P. 2278–2324.
  9. Svertochnaja nejjronnaja set [EHlektronnyjj resurs] / Vikipedija – svobodnaja ehnciklopedija. EHlektron. tekstovye dan. Rezhim dostupa: https://ru.wikipedia.org/wiki/Svertochnaja_nejjronnaja_set, svobodnyjj. – Data obrashhenija: 05.04.2017.
  10. Avedjan EH.D., Galushkin A.I., Selivanov S.A. Sravnitelnyjj analiz struktur polnosvjaznykh i svertochnykh nejjronnykh setejj i ikh algoritmov obuchenija // Informatizacija i svjaz. 2017. № 1. S. 18–30.
  11. Model Zoo [EHlektronnyjj resurs] / BVLC/caffe Wiki – Rezhim dostupa: https://github.com/BVLC/caffe/wiki/Model-Zoo, svobodnyjj. – Data obrashhenija: 06.04.2017.
  12. Osovskijj S. Nejjronnye seti dlja obrabotki informacii: Per. s polsk. I.D. Rudinskogo. M.: Finansy i statistika. 2002. 344 s.

 

June 24, 2020
May 29, 2020

© Издательство «РАДИОТЕХНИКА», 2004-2017            Тел.: (495) 625-9241                   Designed by [SWAP]Studio