Code for Spatial-frequency channels, shape bias, and adversarial robustness

This repository contains code and files required to generate critical band masking images and reproduce experiments from our NeurIPS 2023 (Oral) paper: "Spatial-frequency channels, shape bias, and adversarial robustness". You can find the paper here: https://arxiv.org/abs/2309.13190.

0. Installation

First, clone the repository and install the required python packages into a Conda environment.

# Clone
git clone https://github.com/ajaysub110/critical-band-masking

# Create conda environment
conda create -n cbm_env --python=3.8 pip

# Install requirements
cd critical-band-masking
pip install -r requirements.txt

1. Generating image dataset

For all our critical-band masking experiments on humans and machines, we use a dataset of low-contrast, grayscale ImageNet (Russakovsky et al., 2015) images perturbed with Gaussian noise of 5 possible strengths, bandpass-filtered within 7 possible octave-wide spatial-frequency bands. Filtering code uses pyrtools (Broderick et al., 2019).

You can download the critical-band masking image dataset we used for all human and neural network experiments here. Extract the zip file and save it in the stimuli/ folder.

Otherwise, as shown below you can use scripts in the stimuli/ folder also generate your own critical-band masking image dataset from a dataset of grayscale images in the same directory format as GraySplit.

NOTE: If you intend to reproduce results exactly from the paper, please use the version of CBMSplit in the provided link. It used a different random seed.

A. Download GraySplit

GraySplit is a dataset of 1100 grayscale-converted ImageNet images that we created from the original ImageNet validation dataset. These images are divided into 6 splits, named train, 0, 1, 2, 3, 4, and 5. 0-5 contain 210 images each, and train contains 50, making up a total of 1100 images.

Image files within each split are named according to the following convention: <index>_color_gray_g_<category>.JPEG, where <index> refers to the image's index within the split, and <category> is the object category the image belongs to. As mentioned in the paper, we use categories from 16-class ImageNet (Geirhos et al., 2018) - airplane, bear, bicycle, bird, boat, bottle, car, cat, chair, clock, dog, elephant, keyboard, knife, oven, truck. Each split in GraySplit has a more-or-less equal number of images from each of these categories.

GraySplit forms a source image set for all of our critical band masking experiments in the paper. It can easily be generated from the ImageNet validation set, but we provide a link to GraySplit because its much smaller and more convenient to use as it has been split and sorted by category for you. We apply all of the filtered-noise transformations in the next section to images in GraySplit. The train subdirectory will become the training images (for humans), and the numbered subdirectories, the test images for both humans and networks.

Download GraySplit.zip from here and unzip it to the stimuli/ directory.

B. Generate critical-band masking (CBM) dataset

Now, you can generate our critical-band masking dataset by simple running the stimuli/createCBMDataset.py providing the path to the 16-class ImageNet directory and output directory as arguments.

python stimuli/createCBMDataset.py

The output directory will be organized as follows

outputDirectory
- 0
	... (210 images)
- 1
	... (210 images)
- 2
	... (210 images)
- 3
	... (210 images)
- 4
	... (210 images)
- train
	... (50 images)

All image filenames are formatted as <index>_cbm_noise<noise_sd>_freq<freq_index>_gray_g_<category>.JPEG, where <noise_sd> is the strength of added noise, and <freq_index> is its spatial-frequency. The train folder was used to train human observers in our psychophysics experiments, and was not used for neural network evaluation.

2. Human psychophysics dataset and analysis

The data from our human psychophysics experiment (data), are available in data/human_data.txt. It's in an uncommonly used format, so we provide a demo analysis in the iPython notebook humans/human_analysis.ipynb, where we load the data, compute accuracy heatmaps, thresholds and find the human channel.

3. Neural network evaluation

Finally, we provide code to evaluate neural network models on the critical-band masking task (nn_cbm.py), and also to compute their shape bias (nn_shapebias.py) and adversarial robustness (nn_whitebox.py).

References

1. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., ... & Fei-Fei, L. (2015). Imagenet large scale visual recognition challenge. International journal of computer vision, 115, 211-252.
2. Geirhos, R., Temme, C. R., Rauber, J., Schütt, H. H., Bethge, M., & Wichmann, F. A. (2018). Generalisation in humans and deep neural networks. Advances in neural information processing systems, 31.
3. William Broderick, Pierre-Étienne Fiquet, Zhuo Wang, Zahra Kadkhodaie, Nikhil Parthasarathy, and the Lab for Computational Vision (2019)