Teaching a GAN What Not to Learn

Authors: Siddarth Asokan, Chandra Seelamantula

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The advantage of the reformulation is demonstrated by means of experiments conducted on MNIST, Fashion MNIST, Celeb A, and CIFAR-10 datasets.
Researcher Affiliation Academia Siddarth Asokan Robert Bosch Center for Cyber-Physical Systems Indian Institute of Science Bangalore, India siddartha@iisc.ac.in Chandra Sekhar Seelamantula Department of Electrical Engineering Indian Institute of Science Bangalore, India css@iisc.ac.in
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes We conduct experiments on MNIST [28], Fashion-MNIST [29], Celeb A [30] and CIFAR-10 [31] datasets.
Dataset Splits No The paper describes how positive/negative classes and minority classes are constructed from the datasets, but it does not specify traditional training/validation/test dataset splits with exact percentages, sample counts, or citations to predefined splits for general model training.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies Yes The GAN models are coded in Tensor Flow 2.0 [32].
Experiment Setup Yes In all the cases, latent noise is drawn from a 100-dimensional standard Gaussian N(0100, I100). The ADAM optimizer [34] with learning rate η = 10 4 and exponential decay parameters for the first and second moments β1 = 0.50 and β2 = 0.999 is used for training both the generator and the discriminator. A batch size of 100 is used for all the experiments and all models were trained for 100 epochs, unless stated otherwise.