Wavelet Pooling for Convolutional Neural Networks

Authors: Travis Williams, Robert Li

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on four benchmark classification datasets demonstrate our proposed method outperforms or performs comparatively with methods like max, mean, mixed, and stochastic pooling.
Researcher Affiliation Academia Travis Williams Department of Electrical Engineering North Carolina A&T State University Greensboro, NC 27410, USA tlwilli3@aggies.ncat.edu Robert Li Department of Electrical Engineering North Carolina A&T State University Greensboro, NC 27410, USA eeli@ncat.edu
Pseudocode Yes Figure 4: Wavelet Pooling Forward Propagation Algorithm Figure 5: Wavelet Pooling Backpropagation Algorithm
Open Source Code No The paper does not contain any explicit statement about providing open-source code for the described methodology, nor does it provide a link to a code repository.
Open Datasets Yes We test these methods on benchmark image classification datasets such as Mixed National Institute of Standards and Technology (MNIST) (Lecun et al., 1998), Canadian Institute for Advanced Research (CIFAR-10) (Krizhevsky, 2009), Street House View Numbers (SHVN) (Netzer et al., 2011), and Karolinska Directed Emotional Faces (KDEF) (Lundqvist et al., 1998).
Dataset Splits Yes For the case with dropout, we use the full training set of 73,257 images, a validation set of 30,000 images we extract from the extra training set of 531,131 images, as well as the full testing set of 26,032 images.
Hardware Specification Yes All experiments are run on a 64-bit operating system, with an Intel Core i7-6800k CPU @ 3.40 GHz processor, with 64.0 GB of RAM. We utilize two Ge Force Titan X Pascal GPUs with 12 GB of video memory for all training.
Software Dependencies Yes All CNN experiments use Mat Conv Net (Vedaldi & Lenc, 2015). All simulations in MATLAB R2016b.
Experiment Setup Yes All training uses stochastic gradient descent (Bottou, 2010). For our proposed method, the wavelet basis is the Haar wavelet... All CNN structures except for MNIST use a network loosely based on Zeilers network (Zeiler & Fergus, 2013). We repeat the experiments with Dropout (Srivastava, 2013) and replace Local Response Normalization (Krizhevsky, 2009) with Batch Normalization (Ioffe & Szegedy, 2015) for CIFAR-10 and SHVN (Dropout only)... All pooling methods use a 2x2 window for an even comparison to the proposed method.