我们提出了一种名为 WGAN 的新算法，作为传统 GAN 训练的替代方案。在这个新模型中，我们展示了可以改善学习的稳定性，消除模式崩溃等问题，并提供有意义的用于调试和超参数搜索的学习曲线。此外，我们证明了相应的优化问题是合理的，并提供了广泛的理论工作，突出了与其他分布距离的深层联系。
We introduce a new algorithm named WGAN, an alternative to traditional GAN training. In this new model, we show that we can improve the stability of learning, get rid of problems like mode collapse, and provide meaningful learning curves useful for debugging and hyperparameter searches. Furthermore, we show that the corresponding optimization problem is sound, and provide extensive theoretical work highlighting the deep connections to other distances between distributions.
核心贡献 · Key contributions
提出使用推土机距离的 WGAN，提升训练稳定性。 Proposes WGAN using Earth Mover distance to improve training stability.
展示 WGAN 消除模式崩溃并提供有意义的损失曲线。 Shows WGAN eliminates mode collapse and provides meaningful loss curves.
证明在温和假设下 Wasserstein 损失的连续性和可微性。 Proves continuity and differentiability of Wasserstein loss under mild assumptions.
展示将判别器训练至最优可产生可靠梯度。 Demonstrates training critic to optimality yields reliable gradients.
提供理论分析，连接 EM 距离与其他概率度量。 Provides theoretical analysis linking EM distance to other probability metrics.
在图像生成上实证验证 WGAN，鲁棒性提升。 Empirically validates WGAN on image generation with improved robustness.
局限 · Limitations
权重裁剪实现 Lipschitz 约束粗糙，可能引发问题。 Weight clipping for Lipschitz constraint is crude and may cause issues.
WGAN 对动量优化器（如 Adam）不稳定，需用 RMSProp。 WGAN unstable with momentum optimizers like Adam; requires RMSProp.
损失估计缩放依赖判别器架构，限制跨模型比较。 Loss estimate scaling depends on critic architecture, limiting cross-model comparison.
理论最优性假设判别器无限容量，不切实际。 Theoretical optimality assumes infinite critic capacity, not practical.
实验限于 LSUN-Bedrooms 数据集，泛化性未充分测试。 Experiments limited to LSUN-Bedrooms dataset; generalizability not fully tested.