A REVIEW OF TRANSFER LEARNING: ADVANTAGES, STRATEGIES AND TYPES

Abstract

This study provides an in-depth exploration of Transfer Learning (TL), a powerful machine learning technique that applies knowledge from one domain to enhance learning and performance in a different but related task. Based on the fundamental principle of transferability of experiences, TL emulates human capability to leverage previous knowledge in new tasks. The study discusses the operational mechanism of TL, especially in the context of deep neural networks, where weights of a pretrained model are utilized to initialize a new model. These inherited weights capture the features learned from the source task, subsequently improving the performance in the target task. The concept is further elucidated through the lens of deep convolutional neural networks (CNN), where TL optimizes the training process by reusing the features learned in the earlier layers of a pre-trained model and updating the task-specific last layer for the new task. The paper reviews the diverse application areas of TL, its advantages and disadvantages, as well as its current implementations in scientific literature. The insights presented in this study contribute to the continued development and wider adoption of TL in cutting-edge research and industry applications, owing to its potential to expedite the training process, improve accuracy, and enable better generalization of machine learning models across various disciplines.