Non-parametric transfer learning using intra-domain alignment

Abstract

The main goal of transfer learning is to transfer knowledge from one domain with enough labeled samples to another similar domain but with few labeled samples and even without labels. Unfortunately, the existing methods in the field of transfer learning often require the selection of a pre-trained model, as well as the search and adjustment of many parameters to achieve the desired result, as well as the validation of the selection of the appropriate parameter and model is a very expensive task. And in many applications, it will be impossible due to the lack of labeled data, and this will limit the application of transfer learning in practical applications and will not make significant progress. In this research, we will propose a transfer learning method that, although it creates a competitive advantage compared to the works done in this field, but at the same time, it does not need to adjust and choose the model and validate them. In the ETLS method, we will be able to perform domain adaptation in addition to transfer learning and classification by using intra-domain programming. ETLS is very simple and accurate and at the same time trains the model with very low execution time, the results of experiments on different data sets show that the ETLS method is better than existing traditional methods and neural network methods which need to be implemented several times, it is more accurate and by using it, the performance and classification accuracy of the existing methods can be increased. This simplicity and quick implementation of ETLS model makes it practical in reality and detection devices and sensors.

Keywords: Transfer Learning, Domain Adaptation, Cross-domain Learning, Non-parametric Learning