WGAN with an Infinitely Wide Generator Has No Spurious Stationary Points

Authors: Albert No, Taeho Yoon, Kwon Sehyun, Ernest K Ryu

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Figure 6 presents an experiment with a mixture of 8 Gaussians and Ng = 5, 000. The experiments demonstrate the sufficiency of two-layer networks with random features and that the training does not encounter local minima when Ng is large. Figure 5 visualizes the loss landscape with generator widths Ng = 2 and Ng = 10. For the Ng = 10 case, the parameter space was projected down to a 2D space defined by random directions, as recommended by Li et al. (2018b). We observe the landscape becomes more favorable with larger width.
Researcher Affiliation Academia 1Department of Electronic and Electrical Engineering, Hongik University, Seoul, Korea 2Department of Mathematical Sciences, Seoul National University, Seoul, Korea.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes The code is available at https://github.com/sehyunkwon/Infinite-WGAN.
Open Datasets No Figure 6 presents an experiment with a mixture of 8 Gaussians and Ng = 5, 000. The paper does not provide concrete access information for this dataset.
Dataset Splits No The paper describes experiments with synthetic data (mixture of Gaussians) but does not specify any training, validation, or test dataset splits.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes We fix the maximization stepsize to 1 while letting the minimization stepsize be α > 0. ... Ng = 5, 000, and Nd = 1, 000. ... with generator widths Ng = 2 and Ng = 10.