COMPARISON OF ADAPTIVE MOMENT ESTIMATION OPTIMIZATION AND NESTEROV-ACCELERATED ADAPTIVE MOMENT ESTIMATION OPTIMIZATION ON CONVOLUTIONAL NEURAL NETWORK METHOD FOR FRUIT DETECTION

Abstract

Convolutional Neural Networks are often used in research to conduct training, validation, classification, prediction and detection of images using Deep Neural Network. Optimization algorithm is used to change the hyperparameter values ​​in the Neural Network such as learning rate, optimization is needed to reduce losses and increase the accuracy of the model. Optimization algorithm that is widely used because of its good performance is Adam and Nadam optimization, but the learning rate setting still needs to be updated manually. In this research architecture that was based on VGG16 will be used, Learning Rate Scheduler is used in optimization to control the learning rate value by updating the learning rate value in each step during model training. In this study, a comparison of the optimization of Adam and Nadam was carried out when the Learning Rate Scheduler was used to update the learning rate value in model training and obtained prediction accuracy using Adam 98.85% and Nadam 95.02% and then obtained MAP model performance value using Adam 93.58%. and Nadam 75.28%.