Skew Orthogonal Convolutions
SOC is a convolution layer that has an Orthogonal Jacobian matrix and achieves improved standard and provably robust accuracy over the prior works.
Prerequisites
- Python 3.7+, Pytorch 1.6+
- A recent NVIDIA GPU
How to run?
- Run
python train_robust.py --conv-type skew --block-size BLOCK_SIZE --dataset DATASET_NAME
Demonstration
Citation
@inproceedings{singlafeiziSOC2021,
title={Skew Orthogonal Convolutions},
author={Sahil Singla and Soheil Feizi},
booktitle={ICML},
year={2021}
}