Improving MNIST Image Synthesis via an Optimized Generative Adversarial Network with Transfer Learning and Real-Time Loss Monitoring

Abstract: This paper presents an optimized Generative Adversarial Network (GAN) framework for MNIST image generation using Generative Artificial Intelligence (GenAI). The proposed model leverages the collaborative learning process between a generator and a discriminator network to synthesize realistic handwritten digit images. Training efficiency and model stability are enhanced through dynamic loss monitoring and an optimized generator–discriminator architecture. The benchmark MNIST dataset, consisting of grayscale images of handwritten digits from 0 to 9, is used for training and evaluation. The generator is designed using dense neural network layers to create synthetic images, while the discriminator functions as a binary classifier to distinguish between real and generated samples. Throughout the training process, the losses of both networks are continuously monitored to ensure effective convergence and balanced adversarial learning. The performance of the proposed framework is evaluated using multiple metrics, including accuracy, loss, F1-score, and Receiver Operating Characteristic (ROC) curve analysis. Experimental results demonstrate that the model achieves an average classification accuracy of 90%, indicating its effectiveness in generating high-quality MNIST digit images. Furthermore, this work explores the integration of transfer learning techniques within the GAN framework, providing a foundation for extending similar methodologies to more complex image datasets and real-world applications. Future research may focus on advanced loss functions, improved network architectures, and the application of GAN- based image generation across diverse domains. The proposed framework also serves as a valuable educational and research resource for scholars and practitioners working in the fields of Generative AI and deep learning.

Keywords: Generative AI (GenAI), Generative Adversarial Network (GAN), MNIST, Deep Learning, Discriminator loss, Generator loss