Salient ImageNet: How to discover spurious features in Deep Learning?

Authors: Sahil Singla, Soheil Feizi

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

Reproducibility Variable Result LLM Response
Research Type Experimental We apply our proposed methodology to the Imagenet dataset: we conducted a Mechanical Turk study using 232 classes of Imagenet... For various standard models (Resnet-50, Wide-Resnet-50-2, Efficientnet-b4, Efficientnet-b7), we evaluate their accuracy drops due to corruptions in spurious or core regions...
Researcher Affiliation Academia Sahil Singla & Soheil Feizi University of Maryland, College Park {ssingla,sfeizi}@umd.edu
Pseudocode No The paper does not contain explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes Code and dataset for reproducing all experiments in the paper is available at https://github.com/singlasahil14/salient_imagenet.
Open Datasets Yes We apply our proposed methodology to the Imagenet dataset: we conducted a Mechanical Turk study using 232 classes of Imagenet... Using this methodology, we introduce the Salient Imagenet dataset containing core and spurious masks for a large set of samples from Imagenet... The dataset and anonymized Mechanical Turk study results are also available at the associated github repository.
Dataset Splits No The paper uses pre-trained models and discusses selecting images from the Imagenet training set and validation set, but does not provide explicit training, validation, and test dataset splits for reproducing a model's training process within their framework.
Hardware Specification No The paper does not specify the hardware, such as GPU or CPU models, used to run the experiments.
Software Dependencies No The paper implies the use of Python libraries such as OpenCV and NumPy through code snippets, but it does not specify exact version numbers for any software dependencies.
Experiment Setup Yes We use σ = 0.25 (equation 1)... For optimization, we use gradient ascent with step size = 40, number of iterations = 25 and ρ = 500.