Analysis of the effects of hyperparameters in convolutional neural networks

In this study, literature review and experimental study were carried out on the hyperparameters constituting the structure, working system and network of the irregular neural networks that gained popularity in the definition of the day-to-day picture in 2012 within the scope of IMAGE-NET contest. In the study, ILSVRC2012 training dataset consisting of 50 classes and 600 samples and different option values for convolutional neural network hyperparameters were determined and trainings were conducted on the learning structure of the deep learning client, parameter and evaluation server included in the supercomputers. Consequently, these trainings, the model performances were evaluated through diagrams and charts and new hyper parameter values were created and additional trainings were made. As a result of 410 separate trainings in total, it has been determined that preprocessing of data sets, learning rate selection in accordance with optimizer, packet normalization and use of dropout process, increases model performance.