A0573
Title: A Bayesian non-parametric approach: Integrating VAEs and GANs using Wasserstein and MMD
Authors: Forough Fazeliasl - University of Hong Kong (Hong Kong) [presenting]
Michael Minyi Zhang - University of Hong Kong (Hong Kong)
Abstract: Generative models like GANs and VAEs have shown promise in generating realistic images. While GANs produce sharp images, they often miss out on the full diversity of the target distribution. On the other hand, VAEs generate diverse images but tend to be blurry. To overcome these limitations, a novel approach is proposed that combines GANs and VAEs using a Bayesian non-parametric (BNP) framework. The method incorporates Wasserstein and maximum mean discrepancy (MMD) measures in the loss function to effectively learn the latent space and generate diverse, high-quality samples. By merging the discriminative power of GANs and the reconstruction capabilities of VAEs, our model outperforms existing approaches in tasks like anomaly detection and data augmentation. Additionally, an extra generator is introduced in the code space to explore areas overlooked by the VAE. The BNP perspective allows for the modeling of data distribution using an infinite-dimensional space, providing flexibility and reducing overfitting risks. This framework enhances the performance of GANs and VAEs, resulting in a more robust generative model suitable for various applications. By combining the strengths of these models and mitigating their weaknesses, the approach opens new possibilities for generating high-quality images that closely resemble real ones.
