The Impact of Multi-Optimizers and Data Augmentation on TensorFlow Convolutional Neural Network Performance

Abstract

This paper introduces a new methodology for Alzheimer disease (AD) classification based on TensorFlow Convolu-tional Neural Network (TF-CNN). The network consists of three convolutional layers to extract AD features, a flatten-ing layer to reduce dimensionality, and two fully connected layers to classify the extracted features. The whole purpose of TensorFlow is to have a computational graph. To boost the classification performance, two main con-tributions have been done: data augmentation and multi-optimizers. The data augmentation helps to decrease over-fitting and increase the performance of the model. The training dataset images are augmented by normalizing, rotating, and cropping them. Four different optimizers are used with the TF-CNN, Adagrad, ProximalAdagrad, Adam, and RMSProp to achieve accurate classification. The re-sult demonstrates that the loss value of the Adam and RMSProp optimizers was lower than the Adagrad and ProximalAdagrad optimizers. The classification accuracy using Adam optimizer is 95.8%, while it reaches 100% when using RMSProp optimizer.