Novel semi-supervised learning approach for descriptor generation using artificial neural networks

Abstract

The rise of machine learning and neural networks has opened many doors for making various arduous real-life tasks far more accessible, in addition to their ability to analyze vast amounts of data that are considered to be impossible for humans to process. Neural networks are an essential topic as they can be applied in many real-life applications, such as image, video and sound matching, making them a very attractive research area. Numerous methods and approaches are available for training neural networks, but this paper is concerned with only the semi-supervised training approach, for which a new "enhanced semi-supervised" learning method is proposed. Semi-supervised learning means that machines, such as computers, can learn in the presence of datasets that are both labeled and unlabeled. In contrast, the supervised learning approach can be applicable with labeled data only. A novel semi-supervised learning approach for descriptor generation using artificial neural networks is proposed to control the values that are output by the neural network. However, no interaction with the assignment of these values to each input group occurs, nor is the space where the output values belong utilized. Thus, this method seeks to provide a more efficient learning approach with a more even distribution of the output throughout the output field of space, resulting in a more effective learning approach. The handwritten digit experiment showed an accuracy of 85.27%, while Alzheimer's detection experiment recorded an accuracy of 99.27%. The results after applying the proposed method to two sets of experimental data revealed a significant improvement in accuracy compared with the use of Siamese neural networks in different applications.