Timeline for Why Gaussian Process Regression (GPR) is non-parametric?
Current License: CC BY-SA 4.0
12 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Nov 20 at 8:39 | comment | added | CfourPiO | @JohnMadden Thanks a lot for the comment. Coming from a different background other than statistics, I sometimes expect everything to be the way I understand them. However, I should realize this is research and it evolves over time. Thanks for all the help. | |
| Nov 18 at 18:43 | comment | added | Alexis | +1 Nice! I like the way you mark the distinction. | |
| Nov 15 at 23:47 | comment | added | John Madden | @CfourPiO Regarding: "However, I still believe it is inappropriate to call estimation techniques parametric or non-parametric. " Keep in mind that there never was a committee assembled to determine the best names for statistical concepts: as we were discovering these ideas, names were proposed and some happened to stick, all for purely sociological reasons. Surely there are better ways to name all kinds of ideas. But in order to communicate effectively with the rest of the statistical community, we must nevertheless use the existing terminology in order to be understood. | |
| Nov 15 at 6:54 | comment | added | CfourPiO | I'm accepting this answer as it is consistent with the literature. | |
| Nov 15 at 6:34 | vote | accept | CfourPiO | ||
| Nov 15 at 6:27 | comment | added | CfourPiO | All models are parametric, sure. However, all these approaches/ techniques (GP and KNN too) are also parametric , otherwise there would not be any need to estimate these parameters. Can one fit the data of a GP without estimating the kernel? These kernel define the spectral content, which is physical, not just tuning parameters. | |
| Nov 15 at 6:24 | comment | added | CfourPiO | Intriguing. Thank you for the answer. This definition goes well with what we observe. However, I still believe it is inappropriate to call estimation techniques parametric or non-parametric. I have a problem with the semantics now. The techniques where the data is needed to make a prediction along with the parameters (can be parameters of the kernel), should be called something else but not non-parametric. The data is required, but the underlying process is defined by its frequency content (kernel). I don't know if other people find the semantics a bit absurd, but I definitely do. | |
| Nov 14 at 18:24 | history | edited | John Madden | CC BY-SA 4.0 |
added 1 character in body
|
| Nov 14 at 17:08 | history | edited | John Madden | CC BY-SA 4.0 |
edited body
|
| Nov 14 at 17:02 | history | edited | John Madden | CC BY-SA 4.0 |
edited body
|
| Nov 14 at 16:57 | history | edited | John Madden | CC BY-SA 4.0 |
added 117 characters in body
|
| Nov 14 at 16:50 | history | answered | John Madden | CC BY-SA 4.0 |