Don't let your Discriminator be fooled

Authors: Brady Zhou, Philipp Krähenbühl

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that adversarial robustness both improves the visual quality of the results, as well as stabilizes the training procedure across a wide range of architectures, hyper-parameters and training objectives. We perform all our experiments on the MNIST (Le Cun et al., 1998), CIFAR10 (Krizhevsky, 2009) and Celeb A (Liu et al., 2015) datasets. Table 1 shows the quantitative results.
Researcher Affiliation Academia Brady Zhou Department of Computer Science University of Texas brady.zhou@utexas.edu Philipp Krähenbühl Department of Computer Science University of Texas philkr@cs.utexas.edu
Pseudocode No The paper includes mathematical formulations and theorems but no pseudocode or algorithm blocks.
Open Source Code No We will publish the code and data used to perform our experiments upon acceptance.
Open Datasets Yes We perform all our experiments on the MNIST (Le Cun et al., 1998), CIFAR10 (Krizhevsky, 2009) and Celeb A (Liu et al., 2015) datasets.
Dataset Splits No The paper mentions using MNIST, CIFAR10, and Celeb A datasets for experiments, but it does not specify the training, validation, or test split percentages or methodology.
Hardware Specification No The paper does not specify the hardware used for running the experiments (e.g., GPU/CPU models, memory).
Software Dependencies No The paper mentions using ADAM for optimization, but it does not specify any software dependencies with version numbers (e.g., Python, PyTorch/TensorFlow versions).
Experiment Setup Yes We train using a weight decay term λ = 10 4, with batch size n = 64 and optimize using ADAM (Kingma & Ba, 2014) with h = 2 10 4 and β0 = 0, β1 = 0.9 for 50 epochs on CIFAR10 and 25 epochs on Celeb A. We use a latent vector of dimension z = 128 and use a unit Gaussian for our sampling distribution. All convolutional blocks are replaced with residual blocks (He et al., 2016), the generator employs batch normalization (Ioffe & Szegedy, 2015) and Re LU nonlinearities, while the discriminator uses instance normalization (Ulyanov et al., 2016) and Leaky Re LU.