[Paper Review] Stabilizing GAN Training with Multiple Random Projections

Behnam Neyshabur, Srinadh Bhojanapalli+1 more|arXiv (Cornell University)|May 22, 2017

Image Processing and 3D ReconstructionComputer Science18 references71 citations

TL;DR

The paper proposes training a single generator against an array of discriminators, each viewing a different random low-dimensional projection of the data, to stabilize GAN training and improve sample quality. Experimental results show higher-quality images than traditional single-discriminator GANs.

ABSTRACT

Training generative adversarial networks is unstable in high-dimensions as the true data distribution tends to be concentrated in a small fraction of the ambient space. The discriminator is then quickly able to classify nearly all generated samples as fake, leaving the generator without meaningful gradients and causing it to deteriorate after a point in training. In this work, we propose training a single generator simultaneously against an array of discriminators, each of which looks at a different random low-dimensional projection of the data. Individual discriminators, now provided with restricted views of the input, are unable to reject generated samples perfectly and continue to provide meaningful gradients to the generator throughout training. Meanwhile, the generator learns to produce samples consistent with the full data distribution to satisfy all discriminators simultaneously. We demonstrate the practical utility of this approach experimentally, and show that it is able to produce image samples with higher quality than traditional training with a single discriminator.

Motivation & Objective

• Improve stability of GAN training in high dimensions by reducing discriminator dominance.
• Maintain consistency so the generator learns the full data distribution despite partial views.
• Demonstrate that multiple random projections yield higher-quality generated images than single-discriminator GANs.
• Provide practical implementation guidance for applying random projection discriminators to image data.

Proposed method

• Train a generator G against multiple discriminators {Dk} each operating on a random low-dimensional projection of data via Wk^Tx.
• Use projections realized as convolutional layers with random Gaussian filters to create image-like projected inputs.
• Discriminators maximize their own projection-specific loss; the generator minimizes the average of these losses across all discriminators.
• Formally optimize min_G max_{D1,...,DK} sum_k V(Dk,G) where V(Dk,G) = E_x~Px[log Dk(Wk^T x)] + E_z~Pz[log(1 - Dk(Wk^T G(z)))]
• Employ random projections with m < d to induce information bottlenecks, facilitating stable gradient flows to G.
• Provide theoretical justification that matching marginals along enough projections implies the full joint distribution is matched under smoothness assumptions.

Experimental results

Research questions

• RQ1 Does training a generator against multiple low-dimensional projections provide stronger and more persistent gradients than a single full-view discriminator?
• RQ2 Can a generator learn the full data distribution by satisfying multiple partial-view discriminators?
• RQ3 How does the number of projections K affect training stability and sample quality?
• RQ4 What practical implementation choices (architecture, projections) influence stability and quality in high-dimensional image generation?
• RQ5 Is the approach transferable to datasets beyond faces (e.g., Imagenet canines) and higher resolutions?

Key findings

• Multiple random projections yield more stable generator training with gradients that remain informative throughout training.
• Generators trained with many discriminators produce higher-quality samples than traditional DC-GANs, with improved detail and fewer distortions.
• Increasing the number of discriminators (K) reduces high-frequency artifacts and improves sample quality when projections are chosen appropriately.
• Training time increases with more discriminators, but generated samples achieve competitive or superior quality after longer training.
• Latent-space interpolations between generated faces using the K-discriminator framework produce plausible intermediate faces, indicating meaningful latent structure.
• Experiments on CelebA-like faces (64x64) show clear quality gains; experiments on Imagenet-canine crops (128x128) demonstrate texture and composition realism despite lack of global plausibility.