A novel similarity based unsupervised technique for training convolutional filters
Yükleniyor...
Tarih
2023-05-17
Yazarlar
Dergi Başlığı
Dergi ISSN
Cilt Başlığı
Yayıncı
IEEE
Erişim Hakkı
info:eu-repo/semantics/openAccess
Özet
Achieving satisfactory results with Convolutional Neural Networks (CNNs) depends on how effectively the filters are trained. Conventionally, an appropriate number of filters is carefully selected, the filters are initialized with a proper initialization method and trained with backpropagation over several epochs. This training scheme requires a large labeled dataset, which is costly and time-consuming to obtain. In this study, we propose an unsupervised approach that extracts convolutional filters from a given dataset in a self-organized manner by processing the training set only once without using backpropagation training. The proposed method allows for the extraction of filters from a given dataset in the absence of labels. In contrast to previous studies, we no longer need to select the best number of filters and a suitable filter weight initialization scheme. Applying this method to the MNIST, EMNIST-Digits, Kuzushiji-MNIST, and Fashion-MNIST datasets yields high test performances of 99.19%, 99.39%, 95.03%, and 90.11%, respectively, without applying backpropagation training or using any preprocessed and augmented data.
Açıklama
This work was supported by the Scientific and Technological Research Council of Turkey (TUBITAK) under Grant 118E293.
Anahtar Kelimeler
Backpropagation, Computer architecture, Convolutional neural networks, Feature extraction, Microprocessors, Task analysis, Training, Unsupervised learning, Extraction, Job analysis, Neural networks, Backpropagation training, Initialization methods, Labeled dataset, Training schemes, Unsupervised techniques, Network, Representation, Neocognitron
Kaynak
IEEE Access
WoS Q Değeri
Q2
Scopus Q Değeri
Q1
Cilt
11
Sayı
Künye
Erkoç, T. & Eskil, M. T. (2023). A novel similarity based unsupervised technique for training convolutional filters. IEEE Access, 11, 49393-49408. doi:10.1109/ACCESS.2023.3277253
