Generative Adversarial Forests for Better Conditioned Adversarial Learning

Yan Zuo, Gil Avraham, Tom DrummondarXiv 2018 · arXiv

Abstract

Training Generative Adversarial Networks (GANs) is notoriously unstable. This paper introduces Generative Adversarial Forests (GAF), which replaces the standard linear head of a GAN’s discriminator with a differentiable decision forest. By leveraging the non-linear power of forests, the authors achieve significantly better "conditioning," leading to rock-solid training stability and sharper, more detailed generated images.

Introduction

The "adversarial" nature of GANs involves a zero-sum game between a Generator (creating fakes) and a Discriminator (detecting fakes). While theoretically sound, practical implementations often suffer from mode collapse or ill-conditioning. The authors argue that the root of this instability lies in the Discriminator’s architecture. Most discriminators use a Fully-Connected (FC) linear layer as their final interpreter. However, linear layers struggle to disentangle the highly complex, non-linear data distributions found in real-world images.

The Conditioning Problem

This architectural limitation leads to "ill-conditioned" gradients. Mathematically, the loss surface looks like a narrow, steep-sided canyon (an ellipsoid) rather than a gentle, round bowl (a sphere). In these narrow canyons, optimization steps often overshoot the mark, causing the training to spiral out of control. Forests help "round out" these wells, making the optimization landscape more manageable.

Figure 1: Comparison of conditioning. (a) A well-conditioned spherical well allows for stable convergence. (b) An ill-conditioned elliptical surface leads to optimization instability.

Methodology: Generative Adversarial Forests

The core innovation is the Generative Adversarial Forest (GAF). Instead of a standard FC layer, the authors embed a decision forest directly into the discriminator. To maintain end-to-end trainability, the authors use "soft" routing via sigmoid functions, allowing the entire forest to be trained using backpropagation.

They adopt a "Soft Residual Forest" approach where each tree contributes a residual value, and the final score is the product of these contributions. Unlike previous hybrid models that alternated between training tree leaves and split nodes, GAF allows for simultaneous updates of all parameters, making it much easier to integrate into standard deep learning workflows.

Experimental Results

The authors evaluated GAF on CIFAR-10, Oxford Flowers, and CelebA datasets. A key finding was the dramatic improvement in the condition number of the Hessian. As shown in Figure 5, the "Tree" head maintains a significantly lower condition number compared to the "Linear" head, which remains high throughout training.

Figure 5: Quantitative comparison of (a) Log Loss and (b) Condition Number (Log10) over training iterations for Tree vs Linear architectures.

This lower condition number translates directly to better stability; GAF models remained stable even when the learning rate was increased tenfold—a scenario where standard DCGAN discriminators immediately failed. Qualitatively, GAF-generated images show significantly more fine detail (like wrinkles in skin or textures in petals) compared to DCGAN and WGAN baselines.

To evaluate generalization, the authors introduced the Competitive-GAN Score, which pits one GAN's generator against another's discriminator. GAF models consistently outperformed baselines, demonstrating a superior ability to generalize to unseen real images without overfitting or memorizing the training set.

Conclusion

Generative Adversarial Forests offer a robust solution to the instability of adversarial training by focusing on discriminator conditioning. By replacing linear heads with differentiable forests, GAF "rounds out" the loss surface, making optimization more reliable. Future work may explore the application of these forest-based heads to the generator or larger-scale datasets like ImageNet.

Summary updated

BibTeX
@misc{zuo2018generative,
  author = {Yan Zuo and Gil Avraham and Tom Drummond},
  title = {{Generative Adversarial Forests for Better Conditioned Adversarial Learning}},
  year = {2018},
  eprint = {1805.05185},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1805.05185},
}
Figure