Constructing Orthogonal Convolutions in an Explicit Manner

Authors: Tan Yu, Jun Li, YUNFENG CAI, Ping Li

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on CIFAR-10 and CIFAR-100 demonstrate that the proposed ECO convolution is faster than SOC in evaluation while leading to competitive standard and certified robust accuracies.
Researcher Affiliation Industry Tan Yu, Jun Li, Yunfeng Cai, Ping Li Cognitive Computing Lab Baidu Research 10900 NE 8th St. Bellevue, Washington 98004, USA {tanyu01,lijun12,caiyunfeng,liping11}@baidu.com
Pseudocode Yes Algorithm 1: The proposed efficient orthogonal convolution. ... Algorithm 2: Constructing the mapping matrix I
Open Source Code No The paper states, "It has been implemented in the Paddle Paddle deep learning platform https://www.paddlepaddle.org.cn." However, this refers to the platform used for implementation, not an explicit release of the authors' own source code for the method described in the paper.
Open Datasets Yes Experiments on CIFAR-10 and CIFAR-100 demonstrate that the proposed ECO convolution is faster than SOC in evaluation while leading to competitive standard and certified robust accuracies.
Dataset Splits No The paper mentions training and testing on CIFAR-10 and CIFAR-100 but does not explicitly provide details about a validation dataset split (e.g., percentages or counts) or its use.
Hardware Specification Yes The training is conducted on a single NVIDIA V100 GPU with 32G memory.
Software Dependencies No The paper mentions "Paddle Paddle deep learning platform" but does not specify a version number for this platform or any other software dependencies.
Experiment Setup Yes The training takes 200 epochs. The initial learning rate is 0.1 when the number of convolution layers is larger than 25 and 0.05 otherwise, and it decreases by a factor of 0.1 at the 50th and the 150th epoch. We set the weight decay as 5e-4. ... By default, we set T = 5 in training and T = 10 in testing.