Deep Generative Models for Data Synthesis and Augmentation in Machine Learning

Abstract

The increasing trend in deep generative modelling, which offers data scarcity and diversity solutions in machine learning, is one of the most recent developments in the field. The data-oriented approaches drive the need for the quality and variety of data sets that could be prevented due to privacy issues and limited resources. Generative deep models, basically GANs (Generative Adversarial Networks) and VAEs (Variational Automatic Encoders), appear to be the most reliable approach to the synthesis and augmentation of the data. These models employ deep learning to basically learn all by itself from raw data without anyone teaching it, which is the basis of modern artificial intelligence. Accuracy issues between overfitting and poor generalization emphasize the need for smart solutions to the problem of data shortness. Deep generative modelling works based on the data distribution, which the model learns by itself and enables a realistic sample generator. The study reviews the proficiency and complexity of GANs, VAEs, and WGANs, comparing the WGANs' capabilities with the former two. Techniques of data augmentation, e.g., repositioning, rotation, and adding Gaussian noise to the dataset, will greatly increase the diversity of the data. Regardless of the training time, all models showcase competitive inference performance and, as a result, may be satisfactorily used in real-time operations. The insights obtained shed light on ways to improve machine learning and artificial intelligence through brain data synthesis, model training, and computational efficiency.