A Fourier Perspective on Model Robustness in Computer Vision

Authors: Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, Justin Gilmer

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

Reproducibility Variable Result LLM Response
Research Type Experimental For our experiments on CIFAR-10, we use the Wide Res Net-28-10 architecture [27], and for our experiment on Image Net, we use the Res Net-50 architecture [16]. When we use Gaussin data augmentation, we choose parameter σ = 0.1 for CIFAR-10 and σ = 0.4 for Image Net. All experiments use flip and crop during training.
Researcher Affiliation Collaboration Dong Yin Department of EECS UC Berkeley Berkeley, CA 94720 dongyin@berkeley.edu; Raphael Gontijo Lopes Google Research, Brain team Mountain View, CA 94043 iraphael@google.com; Jonathon Shlens Google Research, Brain team Mountain View, CA 94043 shlens@google.com; Ekin D. Cubuk Google Research, Brain team Mountain View, CA 94043 cubuk@google.com; Justin Gilmer Google Research, Brain team Mountain View, CA 94043 gilmer@google.com
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper states: 'Our experiment is based on the open source implementation of Auto Augment at https://github.com/tensorflow/models/tree/master/research/autoaugment.' This refers to a third-party open-source implementation used by the authors, not the authors' own code for their specific methodology.
Open Datasets Yes For our experiments on CIFAR-10, we use the Wide Res Net-28-10 architecture [27], and for our experiment on Image Net, we use the Res Net-50 architecture [16].; achieves state-of-the-art robustness on the CIFAR-10-C [17] benchmark.; As for the Image Net-C benchmark...
Dataset Splits Yes Given a model and a validation image X, we can generate a perturbed image with Fourier basis noise.; Error rates are averaged over the entire Image Net validation set.; Error rates are averaged over 1000 randomly sampled images from the test set.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or cloud computing instance specifications used for running its experiments.
Software Dependencies No The paper mentions 'tensorflow' indirectly via a GitHub link for Auto Augment, but does not provide specific version numbers for TensorFlow or any other software dependencies, nor does it list multiple key components with their versions.
Experiment Setup Yes When we use Gaussin data augmentation, we choose parameter σ = 0.1 for CIFAR-10 and σ = 0.4 for Image Net. All experiments use flip and crop during training.