DidEye: An Attempt for Defining and Incorporating Visual Intuition for Convolutional Neural Networks DidEye: Bir Görsel Sezgi Tanimlama Denemesi ve Evrişimsel Sinir Aǧlarina Uygulanmasi

Koç R., YARMAN VURAL F. T.

31st IEEE Conference on Signal Processing and Communications Applications, SIU 2023, İstanbul, Türkiye, 5 - 08 Temmuz 2023 identifier identifier

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/siu59756.2023.10223803
  • Basıldığı Şehir: İstanbul
  • Basıldığı Ülke: Türkiye
  • Anahtar Kelimeler: CNN, information loss, intuition, machine learning
  • Orta Doğu Teknik Üniversitesi Adresli: Evet

Özet

Is it possible to formally define one of the important capabilities of human mind, the intuition, in a mathematical sense? Can we use this definition to develop more robust CNN (Convolutional Neural Networks) model? Through this method, can we develop an algorithm that can recognize incomplete images? In this study, we attempt to find partial answers to the above questions. First, we examined how the performance of CNN algorithms decreased by reducing the amount of information in the test set images in a controlled manner. In order to reduce this decrease, we made a mathematical definition of intuition that could enrich the convolutional neural networks. We used this definition and the intuition module we created to improve the filter outputs in the convolution layer. We used MNIST number dataset to measure the performance of this new CNN model, enriched with an intuition module, called, DidEye performance. Experimental results show that the suggested DidEye model is much more robust and provides higher performance compared to the classical CNN model in a test set containing incomplete images.