Generalized Robust Bayesian Committee Machine for Large-scale Gaussian Process Regression
Authors: Haitao Liu, Jianfei Cai, Yi Wang, Yew Soon Ong
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We generate n = 104, 5 104, 105, 5 105 and 106 training points, respectively, in [0, 1], and select n = 0.1n test points randomly in [ 0.2, 1.2]. We use two realistic datasets, kin40k (8D, 104 training points, 3 104 test points) (Seeger et al., 2003) and sarcos (21D, 44484 training points, 4449 test points) (Rasmussen & Williams, 2006), to assess the performance of our approach. Fig. 2 depicts the comparative results of six aggregation models on the toy example. |
| Researcher Affiliation | Collaboration | 1Rolls-Royce@NTU Corporate Lab, Nanyang Technological University, Singapore 637460 2School of Computer Science and Engineering, Nanyang Technological University, Singapore 639798 3Applied Technology Group, Rolls-Royce Singapore, 6 Seletar Aerospace Rise, Singapore 797575 4Data Science and Artiļ¬cial Intelligence Research Center, Nanyang Technological University, Singapore 639798. |
| Pseudocode | No | The paper describes the proposed method in textual form and through equations, but it does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | We release the demo codes in https://github.com/Liu Hai Tao01/GRBCM. |
| Open Datasets | Yes | We use two realistic datasets, kin40k (8D, 104 training points, 3 104 test points) (Seeger et al., 2003) and sarcos (21D, 44484 training points, 4449 test points) (Rasmussen & Williams, 2006)... The song dataset is partitioned into 450000 training points and 65345 test points... the 11D electric dataset that is partitioned into 1.8 million training points and 249280 test points... The datasets and the pre-processing scripts are available in https://people.orie.cornell.edu/andrew/. |
| Dataset Splits | No | The paper mentions generating training and test points for the toy example (e.g., "We generate n = 104...training points...and select n = 0.1n test points"), and specifies training and test sizes for the kin40k, sarcos, song, and electric datasets. However, it does not explicitly mention or detail a validation set or any cross-validation strategy. |
| Hardware Specification | Yes | We implement the aggregations by the GPML toolbox5 using the SE kernel in (1) and the conjugate gradients algorithm with the maximum number of evaluations as 500, and execute the code on a workstation with four 3.70 GHz cores and 16 GB RAM (multi-core computing in Matalb is employed). |
| Software Dependencies | No | The paper mentions using the "GPML toolbox" and "Matlab," and also refers to "GPy" via a GitHub link (https://github.com/Sheffield ML/GPy). However, it does not provide specific version numbers for any of these software components (e.g., GPML toolbox version X, Matlab version Y, GPy version Z). |
| Experiment Setup | Yes | Each expert is assigned with m0 = 500 data points. We implement the aggregations by the GPML toolbox5 using the SE kernel in (1) and the conjugate gradients algorithm with the maximum number of evaluations as 500. Particularly, we choose m = 200, mb = 0.1n and msod = 2500 for kin40k, and m = 300, mb = 0.1n and msod = 3000 for sarcos. Differently, SVI employs the stochastic gradients algorithm with tsg = 1200 iterations. For the song dataset, we use the foregoing disjoint partition to divide it into M = 720 subsets, and use m = 800, mb = 5000 and tsg = 1300 for SVI; for the electric dataset, we divide it into M = 2880 subsets, and use m = 1000, mb = 5000 and tsg = 1500 for SVI. As a result, each expert is assigned with m0 = 625 data points for the aggregations. |