Keskeiset käsitteet
새로운 Scalable WGF 기반 생성 모델인 Semi-dual JKO를 소개합니다.
Tiivistelmä
Abstract:
WGF provides a promising approach for optimization over probability distributions.
Existing WGF models face scalability challenges due to quadratic training complexity.
Introduction:
Generative models learn the underlying distribution of training data.
Various approaches include Energy-based models, Diffusion models, VAEs, Flow models, GANs, Optimal Transport Maps, and WGF.
Background:
WGF investigates minimizing dynamics of probability density.
JKO scheme is used for numerical approximation of WGF.
Optimal Transport-based Generative Modeling:
UOT problem explores cost-minimizing transport maps.
UOT-based generative models leverage semi-dual form for generative modeling.
Limited Scalability of WGF Models:
JKO models have quadratic complexity, limiting scalability.
Method:
S-JKO model is introduced based on semi-dual form of JKO step.
Experiments:
S-JKO outperforms existing models on CIFAR-10 and CelebA-HQ datasets.
Ablation Studies:
S-JKO shows robustness to varying number of JKO steps and step size.
Conclusion:
S-JKO addresses scalability challenges in generative modeling.
Tilastot
이 논문은 CIFAR-10에서 FID 점수가 2.62, CelebA-HQ에서 5.46을 달성했습니다.
Lainaukset
"Our model significantly outperforms existing WGF-based generative models."
"S-JKO achieves FID scores of 2.62 on CIFAR-10 and 5.46 on CelebA-HQ."