Skew Orthogonal Convolutions

Authors: Sahil Singla, Soheil Feizi

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on CIFAR10 and CIFAR-100 show that SOC allows us to train provably Lipschitz, large convolutional neural networks significantly faster than prior works while achieving significant improvements for both standard and certified robust accuracies.
Researcher Affiliation Academia 1Department of Computer Science, University of Maryland, College Park. Correspondence to: Sahil Singla <ssingla@umd.edu>.
Pseudocode Yes Algorithm 1 Skew Orthogonal Convolution
Open Source Code Yes Code is available at https://github.com/singlasahil14/SOC.
Open Datasets Yes Our experiments on CIFAR10 and CIFAR-100 show that SOC allows us to train provably Lipschitz, large convolutional neural networks significantly faster than prior works while achieving significant improvements for both standard and certified robust accuracies.
Dataset Splits No The paper mentions training and evaluating on CIFAR-10 and CIFAR-100 but does not provide specific train/validation/test dataset splits or methodologies like k-fold cross-validation.
Hardware Specification Yes All experiments were performed using 1 NVIDIA Ge Force RTX 2080 Ti GPU.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes All networks were trained for 200 epochs with an initial learning rate 0.1, dropped by a factor of 0.1 after 50 and 150 epochs. We use no weight decay for training with BCOP convolution as it significantly reduces its performance. For training with standard convolution and SOC, we use a weight decay of 10-4.