Instability and Local Minima in GAN Training with Kernel Discriminators

Authors: Evan Becker, Parthe Pandit, Sundeep Rangan, Alyson K. Fletcher

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical simulations are provided that predictably replicate these behaviors. We conduct a simple experiment on low-dimensional, synthetic data to illustrate the behavior that the theory predicts. We also look at the frequency of GAN failure modes as a function of kernel width of the discriminator. Our main observation is that failure most often occurs when the model is in an isolated points regime at small kernel width.
Researcher Affiliation Academia Evan Becker Dept. CS UCLA evbecker@cs.ucla.edu Parthe Pandit HDSI UC, San Diego parthepandit@ucsd.edu Sundeep Rangan Dept. ECE NYU srangan@nyu.edu Alyson K. Fletcher Dept. Statistics UCLA akfletcher@ucla.edu
Pseudocode No The paper includes mathematical equations and descriptions of dynamics, but it does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code is added as supplementary material
Open Datasets No Two simple datasets for the true data are used: A set of Nr = 4 points arranged on uniformly on the unit circle in dimension d = 2; and a set of Nr = 10 points randomly distributed on the unit sphere in dimension d = 10. In both cases, we initialize Ng = Nr generated point as Gaussians with zero mean and E exj 2 = 1. We approximate the RBF discriminator using a random Fourier feature map as in [24].
Dataset Splits No The paper mentions '40000 training steps' and '100 trials', but it does not specify explicit training, validation, or test dataset splits.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU/CPU models or memory specifications.
Software Dependencies No We only used standard existing code packages such as pytorch. The paper mentions 'pytorch' and 'random Fourier feature map' but does not specify version numbers for these software components.
Experiment Setup Yes We set λ = 0.01 and ηd = ηg = 10 3 and use 40000 training steps. Other details are in the Supplementary material.