Leveraging Semi-Supervised Generative Adversarial Networks to Address Data Scarcity Using Decision Boundary Analysis

Abstract

Convolutional Neural Networks (CNNs) are widely regarded as one of the most effective solutions for image classification. However, developing high-performing systems with these models typically requires a substantial number of labeled images, which can be difficult to acquire. In image classification tasks, insufficient data often leads to overfitting, a critical issue for deep learning models like CNNs. In this study, we introduce a novel approach to addressing data scarcity by leveraging semi-supervised classification models based on Generative Adversarial Networks (SGAN). Our approach demonstrates significant improvements in both efficiency and performance, as shown by variations in the evolution of decision boundaries and overall accuracy. The analysis of decision boundaries is crucial, as it provides insights into the model’s ability to generalize and effectively classify new data points. Using the MNIST dataset, we show that our approach (SGAN) outperform CNN methods, even with fewer labeled images. Specifically, we observe that the distance between the images and the decision boundary in our approach is larger than in CNN-based methods, which contributes to greater model stability. Our approach achieves an accuracy of 84%, while the CNN model struggles to exceed 72%.