FALKON: An Optimal Large Scale Kernel Method
Authors: Alessandro Rudi, Luigi Carratino, Lorenzo Rosasco
NeurIPS 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | An extensive experimental analysis on large scale datasets shows that, even with a single machine, FALKON outperforms previous state of the art solutions, which exploit parallel/distributed architectures. |
| Researcher Affiliation | Academia | Alessandro Rudi INRIA Sierra Project-team, Ecole Normale Sup erieure, Paris Luigi Carratino University of Genoa Genova, Italy Lorenzo Rosasco University of Genoa, LCSL, IIT & MIT |
| Pseudocode | Yes | Algorithm 1 MATLAB code for FALKON. It requires O(n Mt + M 3) in time and O(M 2) in memory. See Sect. A and Alg. 2 in the appendixes for the complete algorithm. |
| Open Source Code | Yes | The code necessary to reproduce the following experiments, plus a FALKON version that is able to use the GPU, is available on Git Hub at https://github.com/LCSL/FALKON_paper . |
| Open Datasets | Yes | Million Songs [36] (Table 2, n = 4.6 105, d = 90, regression). [36] Thierry Bertin-Mahieux, Daniel P. W. Ellis, Brian Whitman, and Paul Lamere. The million song dataset. In ISMIR, 2011. ... IMAGENET (Table 3, n = 1.3 106, d = 1536, multiclass classification). We report the top 1 c-err over the validation set of ILSVRC 2012 with a single crop. |
| Dataset Splits | Yes | For datasets which do not have a fixed test set, we set apart 20% of the data for testing. ... We used a Gaussian kernel with diagonal matrix width learned with cross validation on a small validation set, λ = 10 8 and 105 Nystr om centers. |
| Hardware Specification | Yes | Indeed we used a single machine equipped with two Intel Xeon E5-2630 v3, one NVIDIA Tesla K40c and 128 GB of RAM and a basic MATLAB FALKON implementation |
| Software Dependencies | No | The paper mentions using a 'basic MATLAB FALKON implementation' but does not specify a version number for MATLAB or any other software libraries or dependencies used in the experiments. |
| Experiment Setup | Yes | Million Songs: We used a Gaussian kernel with σ = 6, λ = 10 6 and 104 Nystr om centers. ... TIMIT: We used the same preprocessed dataset of [6] and Gaussian Kernel with σ = 15, λ = 10 9 and 105 Nystr om centers. ... YELP: We used a linear kernel with 5 104 Nystr om centers. |