[Noisebridge-discuss] Build advice for a new system / heavy cluster GPU AI processing?

Mike Schachter mike at mindmech.com
Tue Jul 12 22:10:30 UTC 2011

On Tue, Jul 12, 2011 at 6:21 AM, Sai <sai at saizai.com> wrote:
>> Sounds reasonable - are you deailng with spike data, or something
>> like EEG?
> Primate motor cortex spike data together with timings of direction
> cue, go cue, and start of movement. Need to predict movement direction
> based on ±1s from direction cue (before the go).

Cool! Assuming you're using 8 different directions, at some point you
may consider using a number between 1-360 for the direction and
solve a regression problem instead. I wonder if that would hurt/help...

>> libsvm uses "one-vs-one" multi-class classification. That means,
>> per hyperparam combo, for 8 classes it's training something like
>> (8 choose 2) / 2 = 14 independent SVMs to do it. You might want
>> to look into SVM-lite for multi-class classification:
>> http://svmlight.joachims.org/svm_multiclass.html
> Hm. How is it better?
> In particular, would it be better than eg the sigmoid-only
> CUDA-enabled libsvm variants I pointed to in the OP, if I got a
> beefier nVidia card to use?

>From a quick glance at the paper that SVM-lite's multiclass implementation
is based on, it *might* be better for two reasons. First, it doesn't
do one-vs-one
classification so it doesn't need to train 14 SVMs to solve a single 8-class
problem. However, the algorithm might be more complex and take just as
long, if not longer.

Second, the algorithm described in referenced paper is specifically
developed to solve the multi-class problem, whereas the classical SVMs are
only formulated to solve the binary classification problem. I've never used
SVM-lite though, so I'm only speaking hypothetically. Here's the paper:


At the very least it's worth a try.

Good luck!


More information about the Noisebridge-discuss mailing list