Gaussian-Based Pooling for Convolutional Neural Networks

Authors: Takumi Kobayashi

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experimental results on image classification demonstrate that the proposed method favorably improves performance of various CNNs in comparison with the other pooling methods. We apply the proposed pooling methods (Table 1) to various CNNs on image classification tasks; local pooling layers embedded in original CNNs are replaced with our proposed ones. The classification performance is evaluated by error rates (%) on a validation set provided by datasets.
Researcher Affiliation Academia Takumi Kobayashi National Institute of Advanced Industrial Science and Technology (AIST) 1-1-1 Umezono, Tsukuba, Japan takumi.kobayashi@aist.go.jp
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes The code is available at https://github.com/tk1980/Gaussian Pooling.
Open Datasets Yes on the Cifar100 dataset [13] which contains 50,000 training images of 32 32 pixels and 10,000 validation images of 100 object categories. Image Net dataset [5].
Dataset Splits Yes on the Cifar100 dataset [13] which contains 50,000 training images of 32 32 pixels and 10,000 validation images of 100 object categories
Hardware Specification Yes The CNNs are implemented by using Mat Conv Net [26] and trained on NVIDIA Tesla P40 GPU.
Software Dependencies No The paper mentions 'Mat Conv Net [26]' but does not provide specific version numbers for it or any other software dependencies.
Experiment Setup Yes The network is optimized by SGD with a batch size of 100, weight decay of 0.0005, momentum of 0.9 and the learning rate which is initially set to 0.1 and then divided by 10 at the 80th and 120th epochs over 160 training epochs, and all images are pre-processed by standardization (0-mean and 1-std) and for data augmentation, training images are subject to random horizontal flipping and cropping through 4-pixel padding. for Res Net-based models, we apply the batch size of 256 to SGD with momentum of 0.9, weight decay of 0.0001 and the learning rate which starts from 0.1 and is divided by 10 every 30 epochs throughout 100 training epochs.