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

Sai sai at saizai.com
Tue Jul 12 13:21:03 UTC 2011


On Mon, Jul 11, 2011 at 23:22, Mike Schachter <mike at mindmech.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).

> You have to retrain from scratch per hyperparam combo,

Not just hyperparam combo, but nuisance param combos too (for the grid
search). That's the part that kills me.

> 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?

> I'm most familiar with the random forests in R:

Sounds like I have some reading to do. :-)

> Definitely try random forests, they're the hot shit and people will
> eat it up. Comparing things to linear classifiers (or SVMs with a
> linear kernel) is kind of classic.

*nod*

I also just want to learn more on machine learning while I'm at it…

> If you want to never finish your PhD

… but FWIW this is actually bonus work for one class in my MA that's
finishing. ><

(Hence why I said "within a month".)

> keep going and try out neural networks, deep nets (neural
> networks with many pre-trained hidden layers),

… buuut I did say I'd like to learn more stuff. :-P

Also I do think that this could be turned into a legit paper etc.,
which would be fun. ;-)

> And more importantly, there is a python framework called Theano
> which takes care of parallelizing things on the GPU for you:

Ooo. I'm guessing that requires me coding the thing from scratch
though, which is probably outside of what I can plausibly do in the
near term.

Thanks for all the pointers!

- Sai



More information about the Noisebridge-discuss mailing list