Convolutional Neural Networks With Low-rank Regularization

Authors: Cheng Tai, Tong Xiao, Yi Zhang, Xiaogang Wang, Weinan E

ICLR 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluated the proposed method on CIFAR10 and ILSVRC12 datasets for a variety of modern CNNs, including Alex Net, NIN, VGG and Google Net with success. For example, the forward time of VGG16 is reduced by half while the performance is still comparable. Empirical success suggests that low-rank tensor decompositions can be a very useful tool for speeding up large CNNs.
Researcher Affiliation Academia 1The Program in Applied and Computational Mathematics, Princeton University 2Department of Electronic Engineering, The Chinese University of Hong Kong 3Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor
Pseudocode No The paper presents a theorem that describes the algorithm, but it does not use a structured pseudocode block or algorithm environment.
Open Source Code Yes For exact specifications of the parameters, the reader may check https://github.com/chengtaipu/lowrankcnn.
Open Datasets Yes In this section, we evaluate the proposed scheme on the CIFAR-10 and the ILSVRC12 datasets with several CNN models. ... ILSVRC12 (Russakovsky et al., 2015) is a well-known large-scale benchmark dataset for image classification.
Dataset Splits Yes The learning learning rate is initially set to 0.01 and decreases by a factor of 10 every time the validation error stops decreasing. ... We use the single center crop during the testing stage, and evaluate the performance by the top-5 accuracy on the validation set.
Hardware Specification Yes Our results in Table 3 are based on Nvidia Titan GPUs and Torch 7 with cudnn backend.
Software Dependencies Yes Our results in Table 3 are based on Nvidia Titan GPUs and Torch 7 with cudnn backend.
Experiment Setup Yes The batch size is 100. The learning learning rate is initially set to 0.01 and decreases by a factor of 10 every time the validation error stops decreasing. Some models have dropout units with probability 0.25 inserted after every Re LU.