KANs vs MLPs in OT-GAN

The network architecture for image generation needs help in dimensional transformation. In the traditional Optimal Transport GAN (OT-GAN) model, Multi-Layer Perceptrons (MLPs) are used for dimensional transformation. Although MLPs are widely used, they have significant drawbacks and could perform better in accuracy and image processing. We propose a method that improves the network architecture of OT-GAN to enhance its performance in image generation. First, we optimize the structure of the original OT-GAN model, expanding and replacing certain parts of the network architecture, and then replace all MLPs used for dimensional transformation with Kolmogorov-Arnold Networks (KANs). We verify OT-GAN using KANs on the MNIST and CIFAR-10 datasets. The results of the experiments indicate that the network with the KANs is more stable during training and performs better in terms of image quality and generation diversity than the traditional network using MLPs. These findings demonstrate the feasibility and potential of using KANs in image generation.