Kernel density estimate : http://software.intel.com/en-us/forums/showthread.php?t=71058

[csnatarajan]

Hello,

Since I'm clueless about how to reply to my own post, I am trying to put this in a new thread. I apologize for the confusion/redundancy.

Andrey, thanks for the quick response and sorry for the delayed response. The replies are below:

At the moment VSLand VML components of Math Kernel Library does not have functionality tocompute probability density function fora givendataset. We investigate opportunities for the future VSL development, so your inputs relatedto clarification of the aspects of yourquestion would be appreciated.

Doespdfestimationtake significant amount of time in your applications?

Unfortunately, I do not have a estimate of the time yet! the idea was to use this as a part of a bigger project. For eg the datasets I deal with are ~ 2000^3 so the applcn is run in parallel. I wanted to compute the PDF of the gradient fields in each processor and hence I wanted to llok into VKL, VSL.

What is your typical usage model for pdf computation (for example, the dataset is availablein blocks only, andthenext block is used to update the estimate and toconduct the further analysis; or all the data points are immediately available)?

In my usage all the data points are immediately available.

Canyourinput streamcontain missing points or/and outliers?

I am a little confised by this question, I am guessing this is geared towards error estimation/goodness of fit estimation. I believe by that the data does not have outliers/missing points/a long "tail".

Are you interested in processing of one- or multi-dimensional streams?

it is a multidimensional stream => 3d array

Whatcan bethe typical number ofpoints in the input stream?

per proc ~ 256^3.

Whichother estimates besides pdf arerequired in your applications?

None that comes to mind immediately, maybe MLE? Also, It might be very useful to have the option to use different kernels!

Thanks again,

C.S.N

Thanks,
Andrey

[Andrey_N_Intel]

Hello,

Thank you much, I appreciate yourdetailed answers. Ifthere are more questions later can we contact you to discuss the topic in more details?

Best regards,

Andrey

[csnatarajan]

Thanks again Andrey,

You are certainly welcome to contact me with any questions. I am not a real expert on this subject matter (more like a frequent user) but I can definitely try and answer any Q. to the best of my knowledge!

Cheers,

C.S.N