چكيده به لاتين
Abstract:
Today, convolutional neural networks (CNNs), as a group of deep neural networks, have been used
increasingly in speech recognition and image processing applications. In speech recognition, these
networks are used in audio modeling as well as feature extraction from the spectrum of speech signals.
These networks are also used in image processing in applications such as image classification
and image segmentation. One of the most important problems in CNNs is their low robustness
against noise conditions. Noisy data can be handled by data uncertainty in which, the higher level
of noise in a data point, the higher amount of uncertainty is assigned to that data point. In this
thesis, a new method based on neutrosophic theory is presented for improving CNNs. The main
contributions of this research can be summarized in two sections. First, two new definitions of data
uncertainty are proposed for speech and image data in neutrosophic domain. Second, data uncertainty
is integrated with CNN as a new CNN model. Proposed CNN model can handle noisy data
(data with higher uncertainty) which leads to the improvement of CNN in speech and/or image processing.
To the best of our knowledge, this is the first model of CNN which handle noisy data by
considering data uncertainty and can be applied in any type of data. To show the effectiveness of
the proposed model, experiments were performed on speech and image datasets including Aurora2
and CIFAR-10, respectively. Results demonstrated that the proposed CNN model has more accurate
results in comparison with conventional CNN. In Aurora2 dataset, the proposed CNN model
improved the accuracy by 11% and 4% for clean and noisy conditions, respectively. Finally, for
CIFAR-10 dataset in noisy condition, the proposed model outperformed current models by 4%.
Keywords: Convolution Neural Network, Neutrosophic Theory, Uncertainty
