Image generation using generative adversarial networks (GANs) – a powerful AI model architecture for generating images and audio – has suffered from the chronic issue of training instability despite its many advancements.
But a team of researchers led by Minhyeok Lee, an assistant professor from the School of Electrical and Electronics Engineering at Chung-Ang University in South Korea, developed a novel GAN model that enhances training stability and performance by incorporating kernel functions and histogram transformations.
“Imagine teaching an artist to paint landscapes. Consistent guidance may lead them to produce similar scenes, a phenomenon called mode collapse in machine learning. To prevent this, our PMF-GAN model refines the discriminator’s capabilities, penalizing the generator for producing overly similar outputs, thereby promoting diversity,” the professor said.
Their findings were published in the journal Applied Soft Computing in October 2024.
