CircleGAN: Generative Adversarial Learning across Spherical Circles

Woohyeon Shim, Minsu Cho

Overview

CircleGAN introduces a novel discriminator that maps data onto a hypersphere and evaluates "realness" based on proximity to the sphere's great circle (the equator). By utilizing the largest possible area of the hypersphere to represent real data, the model naturally encourages sample diversity and effectively combats mode collapse, achieving state-of-the-art results on benchmarks like CIFAR and TinyImageNet.

The Challenge: Instability and Mode Collapse

Generative Adversarial Networks (GANs) are notoriously difficult to train. Two major hurdles stand in the way: instability and mode collapse.

Instability often stems from unbounded gradients in the discriminator. While recent works like SphereGAN addressed this by mapping samples onto a bounded hypersphere, they typically evaluated "realness" based on a single prototype point. The problem? A single point is a poor representation of the vast diversity found in real-world data. If the discriminator only rewards the generator for being "close to the center," the generator often falls into mode collapse—producing high-quality but repetitive images that fail to capture the full variety of the target dataset.

The CircleGAN Approach

The authors propose CircleGAN, which reimagines the geometry of the discriminator's embedding space. Instead of a single "real" point, they use a spherical circle.

CircleGAN Architecture and Hypersphere Projection
  1. The Great Circle Strategy: The discriminator learns to project samples onto a unit hypersphere. It is trained to push real samples toward the great circle (the "equator") and fake samples toward the poles.
  2. Geometric Scoring: Because the great circle has the largest circumference and area on the sphere, it provides the most "room" for the model to represent different modes of data. The authors define a "realness score" ($s^{real}$) based on how close a sample is to this equator and a "diversifiability score" ($s^{div}$) based on the radius of the circle the sample occupies.
  3. Adaptive Geometry: Unlike previous spherical models that used fixed coordinates, CircleGAN treats the sphere's center ($c$) and the pivotal vector ($p$, which defines the poles) as learnable parameters.
  4. Conditional Hyperspheres: For labeled datasets, the authors extend this by creating a unique hypersphere for every category. This allows the model to maintain high intra-class diversity.
  5. Auxiliary Losses: To keep the geometry stable, they introduce a Center Estimation Loss (to keep the sphere centered on the data) and a Radius Equalization Loss (to ensure the embedding doesn't collapse or explode).

Core Innovations

  • Area Equals Diversity: The core "aha!" moment is the realization that a point has zero area, while a great circle maximizes it. By spreading the "target" for real samples across a circle rather than a point, the discriminator provides a much richer gradient that doesn't force all generated samples to look identical.
  • Better Projections: The authors argue that their use of $\ell_2$-normalization is superior to the Inverse Stereographic Projection (ISP) used in SphereGAN. ISP tends to cluster samples near the North Pole, wasting the hypersphere's capacity, whereas CircleGAN distributes them more evenly.
  • Class-Wise Hyperspheres: In conditional settings, giving each class its own "world" (hypersphere) prevents the features of one class from interfering with the diversity of another.

Performance Highlights

CircleGAN was evaluated on standard computer vision benchmarks including STL10, CIFAR10, CIFAR100, and TinyImageNet. It consistently outperformed strong baselines like SNGAN and SphereGAN across multiple metrics.

Unconditional Generation Results on CIFAR10 and STL10
  • Unconditional CIFAR10: Achieved an Inception Score (IS) of 8.55 and a Frechet Inception Distance (FID) of 12.3, surpassing SphereGAN (8.39 / 17.1).
  • Conditional CIFAR10: The improvement was even more dramatic, reaching an FID of 5.72 compared to the previous state-of-the-art AMGAN at 7.62.
  • Scalability: On TinyImageNet (a challenging dataset with 200 classes at $64 \times 64$ resolution), CircleGAN maintained high diversity (GAN-train score of 30.4) while other classifier-based models like AMGAN failed to train effectively or collapsed.
  • Qualitative Evidence: t-SNE visualizations (noted in the paper) show that CircleGAN’s generated embeddings almost perfectly overlap with the real data distribution, whereas competing models show significant gaps or "clumping."

Limitations and Future Directions

While CircleGAN shows impressive results, there are a few areas for further exploration:

  • Metric Sensitivity: The authors noted that the Inception Score on STL10 is highly sensitive to small changes, suggesting that current metrics might not fully capture the nuances of spherical embeddings.
  • Computational Overhead: Managing a separate center and pivotal vector for every class in a dataset like the full ImageNet (1,000+ classes) could increase memory requirements, though the authors suggest their method is more efficient than multi-discriminator approaches.
  • Higher Resolutions: The paper focuses on resolutions up to $64 \times 64$. Testing the "great circle" logic on high-resolution synthesis (e.g., $1024 \times 1024$) remains an open challenge.

Summary updated

Abstract

We present a novel discriminator for GANs that improves realness and diversity of generated samples by learning a structured hypersphere embedding space using spherical circles. The proposed discriminator learns to populate realistic samples around the longest spherical circle, i.e., a great circle, while pushing unrealistic samples toward the poles perpendicular to the great circle. Since longer circles occupy larger area on the hypersphere, they encourage more diversity in representation learning, and vice versa. Discriminating samples based on their corresponding spherical circles can thus naturally induce diversity to generated samples. We also extend the proposed method for conditional settings with class labels by creating a hypersphere for each category and performing class-wise discrimination and update. In experiments, we validate the effectiveness for both unconditional and conditional generation on standard benchmarks, achieving the state of the art.
BibTeX
@inproceedings{shim2020circlegan,
  author = {Woohyeon Shim and Minsu Cho},
  title = {{CircleGAN: Generative Adversarial Learning across Spherical Circles}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
  url = {https://openreview.net/forum?id=6CZopl63GO},
}
Figure