A review of various semi-supervised learning models with a deep learning and memory approach

Abstract

Based on data types, four learning methods have been presented to extract patterns from data: supervised, semi-supervised, unsupervised, and reinforcement. Regarding machine learning, labeled data are very hard to access, although unlabeled data are usually collected and accessed easily. On the other hand, in most projects, most of the data are unlabeled but some data are labeled. Therefore, semi-supervised learning is more practical and useful for solving most of the problems. Different semi-supervised learning models have been introduced such as iterative learning (self-training), generative models, graph-based methods, and vector-based techniques. In addition, deep neural networks are used to extract data features using a multilayer model. Various models of this method have been presented to deal with semi-supervised data such as deep generative, virtual adversarial, and Ladder models. In semi-supervised learning, labeled data can contribute significantly to accurate pattern extraction. Thus, they can result in better convergence by having greater effects on models. The aim of this paper was to analyze the available models of semi-supervised learning with an approach to deep learning. A research solution for future studies is to benefit from memory to increase such an effect. Memory-based neural networks are new models of neural networks which can be used in this area.