The Impact of Transfer Learning and Pre-trained Models on Model Performance
Nada Khalil Al-Okbi, Amir H. Gandomi, Samila Sighm, Han-Liwa Xuanm, Haming Jia, Shengxiang Zang, Aseel Smerat, Laith Abualigah · 2025
The concept of transfer learning has become a noteworthy one in the current scenarios of machine learning and most especially in the transfer learning in this paper. Considering the resources needed to train new models, one can make more apples-to-apples comparisons across different tasks by using transfer learning instead of training from scratch models. This is the focus of the evaluation this paper is writing about. The development of different communities has brought various tasks where transfer learning could be applied; schemes of both training and application of the transfer are also provided. Among the results of the modeling the authors demonstrate how effective is the transfer learning within the specific tasks and which model should be applied when. The primary results indicate that the most popular models are capable of obtaining high performance while working on a relatively narrow set of approaches. However, a more delicate view is required, as the work should be done in a specific domain or with a specific objective in mind, and low-level details become crucial to the success of the task. This leads to the demand that several guidelines and rules for practical transfer learning should be provided to the community.