RESULTS OF NEURAL DEEP NETWORKS PARAMETER TUNING FOR FASHION MNIST DATASET RECOGNITION

Abstract

A study of convolutional neural network (CNN) models was conducted in order to obtain better recognition quality of the Fashion MNIST DATASET. From the review, it is known that the Fashion MNIST DATASET recognition set is more difficult than MNIST DATASET recognition. The Fashion-MNIST DATASET is recommended for research on different neural network architectures. The best Fashion MNIST DATASET recognition quality results were obtained by convolutional neural network. In this work, the goal was to improve the recognition quality of the Fashion MNIST DATASET by studying different CNN architectures and their parameters. Two consecutive convolutional neural network architectures were selected from those with Fashion MNIST DATASET recognition quality greater than 93%. A study of their architectures and parameters was conducted. The models correspond to the definition of neural deep networks and have different number of layers. Model studies show the influence of batch_size, validation_split, and validation_data parameters on recognition accuracy, as well as location options for the BatchNormalization layer and the activation layer; the effect of the “filters” parameter for the convolutional layer. In addition, two validation sample selection options were used: the first one was from the training dataset (20%) and the second one was the testing dataset. In the calculations, the number of training epochs was equal to 20. In the training process, the issue of preventing overtraining was solved using the analysis of the loss function. TensorFlow, Keras, Python programming language were used. Software modules were developed and implemented in the Google Colab cloud service. As a result of the research, the recognition quality >93% of the Fashion MNIST DATASET declared in the works of other authors was confirmed, and an improved recognition quality of 94.16% was obtained for one of the selected models. The influence of the batch_size parameter on the recognition quality is substantiated, and the batch_size value is chosen according to the best recognition result of the Fashion MNIST DATASET. Increasing the amount of training data has been shown to improve recognition performance when using valid_data==(X_test, X_test_labels) instead of valid_split for training data. The results of a numerical experiment are presented, which confirm the importance and usefulness of applying regularization methods to solve the retraining problem: adjusting the Dropout layers allowed to improve the recognition accuracy.