ParK: Sound and Efficient Kernel Ridge Regression by Feature Space Partitions
Authors: Luigi Carratino, Stefano Vigogna, Daniele Calandriello, Lorenzo Rosasco
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section we study the performance of Par K on some large-scale datasets (n 10^6, 10^7, 10^9). |
| Researcher Affiliation | Collaboration | Luigi Carratino Ma LGa DIBRIS, University of Genova luigi.carratino@dibris.unige.it Stefano Vigogna Ma LGa DIBRIS, University of Genova vigogna@dibris.unige.it Daniele Calandriello Deep Mind Paris dcalandriello@google.com Lorenzo Rosasco Ma LGa DIBRIS, University of Genova IIT, CBMM MIT lrosasco@mit.edu |
| Pseudocode | Yes | Algorithm 1 Par K: Train [...] Algorithm 2 Par K: Predict |
| Open Source Code | No | The paper states that the experiments are 'implemented in python using pytorch and the FALKON library [22]' but does not provide an explicit statement or link for the open-source release of the Par K algorithm's code. |
| Open Datasets | Yes | We perform experiments on the four large-scale datasets TAXI (n 10^9, d = 9, regression), HIGGS (n 10^7, d = 28, classification), AIRLINE (n 10^6, d = 8, regression), AIRLINE-CLS (n 10^6, d = 8, classification) with the same pre-processing and same random train/test split used in [22]. |
| Dataset Splits | No | The paper mentions using a 'random train/test split' and states 'We do not cross validate hyper-parameters of the local estimators of Par K', but does not specify details about validation dataset splits or percentages. |
| Hardware Specification | Yes | The experiments run on a machine with 2 Intel Xeon Silver 4116 CPUs and 1 GPU NVIDIA Titan Xp. The ram of the machine is 256 GB. |
| Software Dependencies | No | The paper mentions 'implemented in python using pytorch and the FALKON library [22]', but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | We do not cross validate hyper-parameters of the local estimators of Par K. Instead we use the same used by FALKON in the paper [22] with the following exeptions: let λ be the global regularization parameters of FALKON and m the number of the Nyström points, the local estimators of Par K use regularization λq = λρ^-1q and mq = mρq as suggested by the theory. [...] The number of centroids used by Par K and D&C-FALK is Q = 32 for all experiments. |