Reproduce “SONA: Learning Conditional, Unconditional, and Mismatching-Aware Discriminator” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuhta Takida, Satoshi Hayakawa, Takashi Shibuya, Masaaki Imaizumi, Naoki Murata, Bac Nguyen, Toshimitsu Uesaka, Chieh-Hsin Lai and 1 more
Researchers from Sony AI and The University of Tokyo developed SONA, a discriminator for conditional Generative Adversarial Networks that employs distinct, orthogonally projected terms to assess sample naturalness and conditional alignment. This design improves image generation quality, achieving an FID of 6.14 on ImageNet 128x128, and incorporates an adaptive weighting mechanism to dynamically balance learning objectives.
Key insights
- Decoupling the assessment of sample naturalness from conditional alignment through an orthogonal projection for the alignment term effectively reduces interference between these two distinct tasks.
- Explicitly incorporating mismatching-aware supervision using a Bradley-Terry model with negative samples significantly enhances the discriminator's ability to enforce strict conditional alignment.
- Dynamically balancing the multiple objectives (unconditional, conditional, and mismatching-aware) with learnable scalar parameters improves training stability and reduces the complexity of hyperparameter tuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.