|
--------- closed -------------- |
|
it sounds like you have a model for your signal so why can't you do some monte carlo simulations to determine this. It can be quite eye opening! Thanks for your response. I have a generative model for how the observations were collected and the observations themselves. Could you point me to a tutorial on how I might do monte carlo simulations to fit the sources?
(Jan 14 '13 at 23:42)
Igor Nikolskiy
I'm not saying use monte carlo to fit the sources. I am saying use monte carlo to determine the error distribution as you vary your hyperparameters.
(Jan 15 '13 at 08:04)
SeanV
|
|
It's not really my area, but you might try to look into Independent Component Analysis. It's basically something similar. Andrew Ng has it well explained in his set of notes here. Possibly looking further into papers and surveys you can get a cool start. Yep I saw that video lecture a little while ago and at the time Prof. Ng seemed to say that ICA is still in its early days. But that was in 2007, I am wondering where we are now and if people would trust ICA for my problem.
(Nov 17 '11 at 11:39)
Igor Nikolskiy
There's a cool demo here you might like http://research.ics.tkk.fi/ica/cocktail/cocktail_en.cgi
(Nov 19 '11 at 11:18)
Leon Palafox ♦
Just a follow up question then: ICA seems to be an unsupervised approach, but I thought of my problem as supervised because I know the profile of the items that I am looking for. An analogous problem would be to identify objects in google street view, given a training set of those objects in isolation. Would you use ICA for the google street view problem or something else?
(Nov 19 '11 at 14:03)
Igor Nikolskiy
@Igor Nikolsky, Yes ICA is an unsupervised approach, and it is great for extracting instantaneous mixtures of signals, especially ones that are sparse. (That is, have supergaussian PDFs). More generally, it works for signals whos PDFs are not gaussian. I am not sure about your question exactly to help you with ICA though.
(Dec 02 '12 at 23:59)
Tarantulus
|
