Re: Training a neural net (or other classifier) across StarCluster

From: Jacob Barhak <no email>
Date: Thu, 12 Dec 2013 03:04:52 -0600

Hi Alessandro,

The advantage of StarCluster is the ability to run things in parallel. In
the case of a Neural Net work you can have better/faster cross validation
by training a different Neural nets on different nodes.

I know that there are Python Neural net libraries in the Python Scientific
distributions. If you choose an AMI with a scientific Python distribution
installed then this simplifies installations. I am note sure about AMIs
with R, yet I am using an Anaconda AMI which has many scientific python
libraries installed in it.

I hope you find this of help.

                  Jacob






On Wed, Dec 11, 2013 at 1:59 PM, Alessandro Gagliardi <
alessandro.gagliardi_at_glassdoor.com> wrote:

> I would like to train a neural net (or similar classifier) to predict
> one probabilistic value from 9 principal components. When I do it in R
> (using nnet) it caps at a few hundred observations, but that seems too
> small a sample when I have over 40k cases. I know that the space of machine
> learning algorithms (distributed and otherwise) is vast and so was
> wondering if there was something that the StarCluster community might
> recommend. (Ideally it would be something that I could set up on
> StarCluster with minimal difficulty.)
>
> Thanks in advance,
> -Alessandro Gagliardi
>
> _______________________________________________
> StarCluster mailing list
> StarCluster_at_mit.edu
> http://mailman.mit.edu/mailman/listinfo/starcluster
>
>
Received on Thu Dec 12 2013 - 04:04:54 EST