GAN Model
Generative Adversarial Networks (GANs) are a class of deep learning models designed to generate new data instances that resemble a training dataset. Current research focuses on improving GAN training stability, addressing issues like mode collapse, and enhancing controllability over generated outputs, often through integration with other models like diffusion models or reinforcement learning. Applications span diverse fields, including image generation and editing, drug discovery, and data augmentation for tasks where real data is scarce or expensive to obtain, significantly impacting various scientific domains and practical applications. Recent work also highlights the exploration of alternative training methods to improve efficiency and quality, moving beyond traditional adversarial training.
Papers
Differentiable Gaussianization Layers for Inverse Problems Regularized by Deep Generative Models
Dongzhuo Li
Permutation Equivariant Generative Adversarial Networks for Graphs
Yoann Boget, Magda Gregorova, Alexandros Kalousis
Physics guided deep learning generative models for crystal materials discovery
Yong Zhao, Edirisuriya MD Siriwardane, Jianjun Hu
A Generic Approach for Enhancing GANs by Regularized Latent Optimization
Yufan Zhou, Chunyuan Li, Changyou Chen, Jinhui Xu