Initialization of weights with customized values in neural networks and investigation of their effects on classification results

Determining the initial values of convolutional layers in convolutional neural network (CNN) models is a critical issue that directly impacts the learning process and ultimate performance of the network. Incorrectly set initial values can lead to problems such as gradient explosion or vanishing, hindering the training process. Therefore, developing correct weight initialization strategies is essential for enhancing the effectiveness and efficiency of deep learning models. In this thesis, the advantages and disadvantages of the methods used to determine the initial values of convolutional layers in CNNs are compared. In addition to commonly used methods such as Xavier, He, fixed value, and zero initialization, innovative approaches like Gabor filters are also considered. Due to their ability to mimic the biological visual system, Gabor filters have the potential to accelerate the learning process and improve the generalization performance of the network when used as weight initialization. The effects of these methods on performance have been tested using the widely recognized FashionMNIST dataset and the IITD Touchless Palmprint database. In the experimental studies, simple network models were used to observe the direct effects of weight initialization methods on performance. This approach allows for practical observation and comparison of the theoretical advantages of different initialization strategies. The use of common parameters increases the comparability of weight initialization methods and facilitates understanding which method performs better under specific conditions. This methodology aims to enhance the reliability and validity of the study's results, providing robust and replicable findings to the literature.