[Paper Review] OOGAN: Disentangling GAN with One-Hot Sampling and Orthogonal Regularization

Bingchen Liu, Yizhe Zhu+3 more|arXiv (Cornell University)|May 26, 2019

Digital Media Forensic Detection31 references4 citations

TL;DR

OOGAN proposes a novel GAN-based framework for unsupervised disentanglement learning by using alternating one-hot sampling and orthogonal regularization to explicitly disentangle semantic factors in high-resolution images. It achieves superior disentanglement and image quality compared to VAEs and InfoGAN, demonstrating robustness and interpretability through structural innovation rather than loss function tuning.

ABSTRACT

Exploring the potential of GANs for unsupervised disentanglement learning, this paper proposes a novel GAN-based disentanglement framework with One-Hot Sampling and Orthogonal Regularization (OOGAN). While previous works mostly attempt to tackle disentanglement learning through VAE and seek to implicitly minimize the Total Correlation (TC) objective with various sorts of approximation methods, we show that GANs have a natural advantage in disentangling with an alternating latent variable (noise) sampling method that is straightforward and robust. Furthermore, we provide a brand-new perspective on designing the structure of the generator and discriminator, demonstrating that a minor structural change and an orthogonal regularization on model weights entails an improved disentanglement. Instead of experimenting on simple toy datasets, we conduct experiments on higher-resolution images and show that OOGAN greatly pushes the boundary of unsupervised disentanglement.

Motivation & Objective

• To address the limitations of VAE-based disentanglement methods, which suffer from poor generative quality and implicit TC approximation on high-resolution images.
• To exploit the inherent structural advantages of GANs—specifically, active noise sampling—for explicit disentanglement learning without relying on implicit variational inference.
• To introduce a robust, architecture-driven approach to disentanglement that avoids unstable loss-weight tuning and instead uses sampling and regularization strategies.
• To resolve the 'compe-te and conflict' issue in InfoGAN-like models through a structural modification to the generator and discriminator.
• To propose a new, intuitive metric for evaluating disentanglement in the generative component of GANs, focusing on attribute prediction and perceptual diversity.

Proposed method

• Introduces an alternating one-hot sampling procedure where noise vectors are sampled from one-hot distributions to enforce exclusivity and encourage each latent dimension to learn a single semantic factor.
• Applies orthogonal regularization to the generator and discriminator weights to stabilize training and improve disentanglement by reducing feature redundancy.
• Replaces the standard generator and discriminator architecture in InfoGAN with a modified structure that eliminates the 'compe-te and conflict' issue, where multiple dimensions compete to represent the same factor.
• Uses a deterministic or probabilistic head $Q$ to predict latent factors $c$ from generated images, with loss functions including L1 for reconstruction and cross-entropy for one-hot targets.
• Employs a perceptual diversity metric based on a fine-tuned VGG network on CelebA to evaluate disentanglement quality, sensitive to facial attributes.
• Estimates Total Correlation (TC) during training using the method from Higgins et al. (2017), showing that OOGAN maintains consistently low TC values.

Experimental results

Research questions

• RQ1 Can GANs achieve explicit, robust disentanglement through a structured sampling strategy rather than loss function engineering?
• RQ2 Does one-hot sampling during training lead to better disentanglement and higher-quality image generation than standard Gaussian noise in GANs?
• RQ3 How does orthogonal regularization of model weights affect disentanglement and training stability in GANs?
• RQ4 Can architectural modifications resolve the 'compe-te and conflict' issue in InfoGAN-like models, where multiple latent dimensions encode the same factor?
• RQ5 Can a new, compact metric effectively evaluate disentanglement in the generative part of GANs, independent of inference networks?

Key findings

• OOGAN achieves significantly better disentanglement than InfoGAN and VAE-based models on high-resolution CelebA images, especially when the number of latent factors exceeds 6, where InfoGAN fails.
• The one-hot sampling strategy reduces L1 reconstruction loss on one-hot factors by 40-50% compared to InfoGAN, indicating improved disentanglement and predictive accuracy of the latent code.
• OOGAN maintains consistently low Total Correlation (TC) throughout training, while InfoGAN’s TC remains high, confirming better inter-dimensional independence.
• Orthogonal regularization reduces the average cosine similarity among model weights, indicating reduced redundancy and improved disentanglement.
• The proposed perceptual diversity metric shows that OOGAN outperforms baselines in capturing attribute-specific variations, with higher classification accuracy on binary attribute prediction.
• Abllation studies confirm that one-hot sampling contributes most to performance, followed by orthogonal regularization and the compete-free generator structure.