8-BIT OPTIMIZERS IN DEEP LEARNING MODEL TRAINING FOR IMAGE CLASSIFICATION

• Main Author: Gondokusumo, Samuel
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/75909
• Institution: Institut Teknologi Bandung

This research focuses on the implementation of 8-bit optimizers to support the process of training machine learning models, specifically deep learning models for image classification. This optimizer has been chosen because of its novelty, hence it has not been widely tested for various use cases. Tests have been done for the use of 8-bit optimizers in natural language processing (NLP) with impeccable performance compared to previous existing optimizers, henceforth this research will test its performance on the case of image classification. This research is done by comparing the performance of deep learning models utilizing the Adam optimizer, 8-bit optimizers, and no optimizers. With GPU memory usage, model accuracy, and model training time as this research’s metrics, 8-bit optimizers have shown to have the best accuracy and save a significant amount of GPU memory, but with a slightly longer training time. It is hoped that this research is able to serve as a stepping stone for better understanding of optimizers, especially in the deep learning field.