AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

The Unusual Effectiveness of Averaging in GAN Training

Yasin Yaz{\i}c{\i}, Chuan-Sheng Foo, Stefan Winkler, Kim-Hui Yap, Georgios Piliouras, Vijay Chandrasekhar

ICLR 2019 · Architectures and efficiency › RNNs and normalization · approach: Efficient network compression and training · 102 citations (OpenAlex)

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Apply EMA to generator parametersConventional GAN · CIFAR-10IS: 8.14 ± 0.01 -> 8.89 ± 0.12Improved
Apply EMA to generator parametersConventional GAN · CIFAR-10FID: 16.84 ± 1.00 -> 12.56 ± 0.17Improved
Apply EMA to generator parametersConventional GAN · STL-10FID: 22.33 ± 1.59 -> 19.64 ± 1.38Improved
Apply EMA to generator parametersConventional GAN · ImageNetIS: 8.33 ± 0.10 -> 8.88 ± 0.14Improved
Add EMA to the base GANBaseline GAN · Mixture of GaussiansWasserstein-1 distance: 0.0419 ± 0.0054 -> 0.0251 ± 0.0026Improved
Add EMA to Optimistic AdamOptimistic Adam · Mixture of GaussiansWasserstein-1 distance: 0.0494 ± 0.0041 -> 0.0431 ± 0.0042Improved
Add EMA to Consensus OptimizationConsensus Optimization · Mixture of GaussiansWasserstein-1 distance: 0.0286 ± 0.0034 -> 0.0252 ± 0.0035Improved
Add EMA to Zero-GPZero-GP · Mixture of GaussiansWasserstein-1 distance: 0.0363 ± 0.0018 -> 0.0246 ± 0.0068Improved
Use uniform moving average instead of no averagingBaseline GAN · Mixture of GaussiansWasserstein-1 distance: 0.0419 ± 0.0054 -> 0.0274 ± 0.0016Improved
Use EMA discount β=0.9999 rather than β=0.9 (ablation)Conventional GAN · CIFAR-10FID: 17.82 ± 0.43 -> 12.91 ± 0.49Improved
Set parameterized online averaging α=1 (ablation)Conventional GAN · CIFAR-10IS: 8.02 ± 0.20 -> failedHurt
Set parameterized online averaging α=10 (ablation)Conventional GAN · CIFAR-10FID: 17.32 ± 0.53 -> 14.55 ± 1.01Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy