Deep Semi-Random Features for Nonlinear Function Approximation

Authors: Kenji Kawaguchi, Bo Xie, Le Song

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

Reproducibility Variable Result LLM Response
Research Type Experimental We compare semi-random features with random features (RF) and neural networks with Re LU on both UCI datasets and image classification benchmarks.
Researcher Affiliation Academia Kenji Kawaguchi Massachusetts Institute of Technology Bo Xie, Le Song Georgia Institute of Technology
Pseudocode No The paper describes the mathematical formulations of the models but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes The source code of the proposed method is publicly available at: http://github.com/zixu1986/semi-random.
Open Datasets Yes We compare semi-random features with random features (RF) and neural networks with Re LU on both UCI datasets and image classification benchmarks. [...] MNIST is a popular dataset for recognizing handwritten digits. [...] CIFAR 10 contains internet images [...] The Street View House Numbers (SVHN) dataset contains house digits collected by Google Street View.
Dataset Splits No The paper provides training and test set sizes (e.g., "60,000 for training and 10,000 for test" for MNIST), but does not explicitly mention a separate validation split or cross-validation strategy in the main text for reproduction.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., GPU models, CPU types).
Software Dependencies No The paper mentions using "tensorflow" for experiments but does not specify its version or any other software dependencies with version numbers.
Experiment Setup Yes The network architecture used on this dataset is multi-layer networks with l = [1, 2, 4] hidden layers and k = [1, 2, 4, 8, 16] d hidden units per layer where d is the input data dimension. [...] We use a convolution neural network consisting of two convolution layers, with 5 5 filters and the number of channels is 32 and 64, respectively. Each convolution is followed by a maxpooling layer, then finally a fully-connected layer of 512 units with 0.5 dropout.