Conceitos Básicos
The author integrates various OT-based GANs, emphasizing the importance of strictly convex functions and the cost function in enhancing training stability and preventing mode collapse. Additionally, a novel method is proposed to address the τ-sensitivity of UOTM while improving performance.
Resumo
The content delves into the significance of optimal transport theory in generative modeling, focusing on the role of convex functions and cost functions in stabilizing training dynamics. A novel approach is introduced to enhance robustness and performance in UOTM models.
- Optimal Transport (OT) theory bridges distributions with minimal cost.
- OT-based generative models benefit from adversarial training objectives.
- Convex functions stabilize training dynamics and prevent mode collapse.
- The proposed UOTM-SD method gradually adjusts divergence terms for improved performance.
Estatísticas
Our approach achieves a FID score of 2.51 on CIFAR-10 and 5.99 on CelebA-HQ-256.
The precision metric results show improvements from WGAN to OTM/UOTM.
The recall metric highlights better mode coverage by UOTM compared to UOTM w/o cost.
Citações
"Setting g1 and g2 as strictly convex functions significantly enhances training stability."
"The cost function plays a crucial role in preventing mode collapse in OT-based GANs."