[论文解读] Mode Regularized Generative Adversarial Networks

Tong Che, Yanran Li+3 more|arXiv (Cornell University)|Dec 7, 2016

Generative Adversarial Networks and Image Synthesis参考文献 25被引用 229

一句话总结

本文提出基于自编码器的正则化项和两步 MDGAN 训练 regime 来稳定 GAN 并缓解缺失模态，从而提升多样性与样本质量。

ABSTRACT

Although Generative Adversarial Networks achieve state-of-the-art results on a variety of generative tasks, they are regarded as highly unstable and prone to miss modes. We argue that these bad behaviors of GANs are due to the very particular functional shape of the trained discriminators in high dimensional spaces, which can easily make training stuck or push probability mass in the wrong direction, towards that of higher concentration than that of the data generating distribution. We introduce several ways of regularizing the objective, which can dramatically stabilize the training of GAN models. We also show that our regularizers can help the fair distribution of probability mass across the modes of the data generating distribution, during the early phases of training and thus providing a unified solution to the missing modes problem.

研究动机与目标

• Motivate GAN training instability and mode missing problems due to discriminator shape in high-dimensional spaces.
• Propose regularizers that add stable geometric and distributional signals to GAN objectives.
• Encourage the generator to align the data and generation manifolds and promote fair mass distribution across modes.
• Develop a two-stage training procedure (manifold step and diffusion step) to improve stability and sample quality.

提出的方法

• Introduce a geometric metric regularizer that minimizes distance between generated samples and their reconstructions via an encoder, using d(x, G(E(x))).
• Incorporate a mode regularizer that uses log D(G(E(x))) to encourage exploration of nearby data modes.
• Propose a two-step Manifold-Diffusion training (MDGAN) with a manifold step to align manifolds and a diffusion step to distribute mass across modes.
• Define a regularized generator objective T_G that combines the adversarial term with reconstruction and mode-matching terms.
• Define a regularized encoder objective T_E that includes reconstruction and adversarial alignment terms.
• Discuss evaluation metrics for mode coverage, including inception/MODE scores and a third-party discriminator-based missing mode estimator.

实验结果

研究问题

• RQ1 How do regularizers affect stability of GAN training and gradient behavior?
• RQ2 Can geometry-based and reconstruction-based signals reduce missing modes and encourage fair mode representation?
• RQ3 Does a two-step manifold-diffusion training scheme improve both sample quality and mode diversity?
• RQ4 What metrics best capture mode coverage and sample quality in regularized GANs?
• RQ5 How do proposed regularizers compare to standard GANs on tasks with multi-modal data distributions?

主要发现

• Regularizers dramatically stabilize GAN training and reduce model variance.
• The regularizers help distribute probability mass more fairly across data modes, addressing missing modes.
• MDGAN improves sample coherence and sharpness compared to baseline GAN variants on CelebA.
• Quantitative metrics show improved MODE scores and reduced missing modes on compositional MNIST experiments.
• Qualitative samples indicate higher visual fidelity and diverse modes with the proposed approach.
• Compared to several baselines, the MDGAN/Regularized-GAN approaches achieve better balance between diversity and quality.