Spatial-frequency channels, shape bias, and adversarial robustness

Authors: Ajay Subramanian, Elena Sizikova, Najib Majaj, Denis Pelli

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Here, we introduce critical band masking as a task for network-human comparison and test 14 humans and 76 neural networks on 16-way Image Net categorization in the presence of narrowband noise. We find that humans recognize objects in natural images using the same one-octave-wide channel that they use for letters and gratings, making it a canonical feature of human object recognition. Unlike humans, the neural network channel is very broad, 2-4 times wider than the human channel.
Researcher Affiliation Academia Ajay Subramanian New York University as15003@nyu.edu Elena Sizikova New York University es5223@nyu.edu Najib J. Majaj New York University najib.majaj@nyu.edu Denis G. Pelli New York University denis.pelli@nyu.edu
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Our human and network datasets as well as code required to reproduce our experiments are available publicly at https://github.com/ajaysub110/critical-band-masking.
Open Datasets Yes In all our experiments, human observers recognized objects in images from Image Net [5], a popular dataset for neural network analysis.
Dataset Splits Yes We tested 76 pretrained neural networks on the same task that we used to test humans: 16-way categorization of noise-masked images from our critical band masking dataset...Additionally, we also measured the accuracy of all networks on 1000-way classification of all 50,000 available unperturbed Image Net validation set images.
Hardware Specification No NYU s High Performance Computing (HPC) cluster for enabling our neural network experiments and analysis.
Software Dependencies No The paper mentions 'Py Torch’s torchvision library', 'Lab JS', 'JATOS', and 'Adversarial Robustness Toolbox' but does not specify their version numbers.
Experiment Setup Yes Grayscale Image Net images were converted to low-contrast (20%) and perturbed with Gaussian noise that was filtered within various spatial-frequency bands. ... To the preprocessed images, we added Gaussian noise of 5 strengths (standard deviations) (0.0, 0.02, 0.04, 0.08, 0.16), filtered into 7 octave-wide (doubling of frequency) spatial-frequency bands (centered at 1.75, 3.5, 7.0, 14.0, 28.0, 56.0, 128.0 cycles/image)... Images were first resized to 256 256 and center-cropped to 224 224... Images were normalized to Image Net pixel mean and variance... generated adversarial perturbations for 1000 images (each belonging to a different category) from the Image Net validation dataset using projected gradient descent (PGD; L = 0.1, max. iterations = 32).