Convolutional Kernel Networks
Authors: Julien Mairal, Piotr Koniusz, Zaid Harchaoui, Cordelia Schmid
NeurIPS 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our methodology on visual recognition tasks where CNNs have proven to perform well, e.g., digit recognition with the MNIST dataset, and the more challenging CIFAR-10 and STL-10 datasets, where our accuracy is competitive with the state of the art. |
| Researcher Affiliation | Academia | Julien Mairal, Piotr Koniusz, Zaid Harchaoui, and Cordelia Schmid Inria firstname.lastname@inria.fr LEAR team, Inria Grenoble, Laboratoire Jean Kuntzmann, CNRS, Univ. Grenoble Alpes, France. |
| Pseudocode | Yes | Algorithm 1 Convolutional kernel network learning the parameters of the k-th layer. Algorithm 2 Convolutional kernel network computing the k-th map form the (k 1)-th one. |
| Open Source Code | Yes | Open-source code for learning our convolutional kernel networks is available on the first author s webpage. |
| Open Datasets | Yes | We evaluate our methodology on visual recognition tasks where CNNs have proven to perform well, e.g., digit recognition with the MNIST dataset, and the more challenging CIFAR-10 and STL-10 datasets. Yet, we achieve competitive results on several datasets such as MNIST [22], CIFAR-10 [20] and STL-10 [13]. |
| Dataset Splits | Yes | We select the regularization parameter of the SVM by 5-fold cross validation when the training size is smaller than 20 000, or otherwise, we keep 10 0000 examples from the training set for validation. We select the best architectures on a validation set of 10 000 examples from the training set for CIFAR-10, and by 5-fold cross-validation on STL-10. |
| Hardware Specification | No | The paper mentions the use of “GPUs” for training but does not provide specific models or other hardware specifications. |
| Software Dependencies | No | The paper mentions using “Matlab and an L-BFGS-B solver [9] interfaced by Stephen Becker” and “Lib Linear [16]” but does not provide specific version numbers for any of these software dependencies. |
| Experiment Setup | Yes | The patches P k are typically small; we tried the sizes m m with m = 3, 4, 5 for the first layer, and m = 2, 3 for the upper ones. The number of filters pk in our experiments is in the set {50, 100, 200, 400, 800}. The downsampling factor γk is always chosen to be 2 between two consecutive layers, whereas the last layer is downsampled to produce final maps ξk of a small size say, 5 5 or 4 4. For the gradient map ϕ0, we approximate the Gaussian kernel e(1/σ2 1) ϕ0(z) ϕ 0(z ) H0 by uniformly sampling p1 = 12 orientations, setting σ1 = 2π/p1. Finally, we also use a small offset ε to prevent numerical instabilities in the normalization steps ψ(z) = ψ(z)/ max( ψ(z) 2, ε). We always run the L-BFGS-B algorithm for 4 000 iterations, which seems to ensure convergence to a stationary point. The regularization parameter of the SVM is always selected on a validation set or by 5-fold cross-validation in the range 2i, i = 15 . . . , 15. |