How Convolutional Neural Networks Deal with Aliasing

Python scripts for reproducing the results from the paper: "How Convolutional Neural Networks Deal with Aliasing".

Antônio H. Ribeiro and Thomas B. Schön "How Convolutional Neural Networks Deal with Aliasing". IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021.

@inproceedings{ribeiro_how_2021,
author={Ant\^onio H. Ribeiro and Thomas B. Sch\"on},
title={How Convolutional Neural Networks Deal with
Aliasing},
year={2021},
publisher={IEEE},
booktitle={2021 {IEEE} International Conference on Acoustics, Speech and Signal Processing, {ICASSP}}

Preprint: https://arxiv.org/abs/2102.07757

---

The folders in this repository contain two experiments:

1. classifying-oscillations: toy example designed to assess the ability of convolutional neural networks to resolve between different frequencies at its input.
2. quantifying-aliasing: Scripts for quantifying to what extent aliasing takes place in the intermediate layers of the neural network

Requirements

The file requirements.txt contains the python modules required. The versions specified are the ones the code has been tested on. Nonetheless, I believe lower versions of most packages should also work. One exception is matplotlib where I observed that using versions different than the 3.2.1 might yield minor changes (namely, different axis ticks).

Finally, some experiments also require ImageNet validation set. I include basic instructions for applying for the license, downloading the dataset and extracting it here.