Controlled Ablation of United Loss

Determine the causal contribution of the United Loss consensus mechanism to training stability and final PSNR in the current v4 loss-guided multi-expert GAN architecture by conducting controlled experiments with and without United Loss and across varied consensus weights.

Background

The paper presents United Loss as a 10% consensus regularizer intended to stabilize the three-discriminator training process. Although the reported training behavior is consistent with this mechanism improving convergence and preserving branch specialization, the current v4 architecture has not undergone a controlled comparison isolating United Loss from the other architectural and training components.

The proposed unresolved study requires comparing the completed United Loss configuration against a configuration with the consensus weight set to zero, as well as testing different consensus strengths. The authors hypothesize that removing United Loss would increase training oscillation, potentially homogenize the expert branches, and reduce final PSNR, but these effects remain unverified on the current architecture because each experiment requires a separate two-to-three-month training run.

References

A controlled ablation of United Loss on the current v4 architecture has not been completed (requiring a separate 2--3 month training run), we report observational evidence consistent with the hypothesized consensus mechanism.

— Sign Language Video Synthesis via Loss-Guided Multi-Expert GANs  (2608.13368 - Nong et al., 13 Aug 2026) in Section 4.3, “Planned Ablation: United Loss Contribution”; Section 5, “Limitations”