I am working on continuous data(speech), and would like to build a deep belief network for classification purpose.

I started with the matrbm library, but it doesn't have real-valued nodes. Is there any good library that allows to create DBNs with GB-RBMs?

Also, I would like to get some idea on the physical interpretation of the model.

asked Mar 04 '13 at 19:26

Niharjyoti%20Sarangi's gravatar image

Niharjyoti Sarangi
1222

1

Integrate Charlie Tang's matlab code for Gaussian RBMs into your stuff: http://www.cs.toronto.edu/~tang/code/GaussianRBM.m

(Mar 05 '13 at 08:49) osdf

Thanks @osdf . Can you help me with a little more detail on how it works?

(Mar 06 '13 at 03:57) Niharjyoti Sarangi

What do you want to know exactly? It is a GRBM that learns also the (conditional) variance for every visible unit (though the learned variance is independent of the setting of the hidden units). Details are described in http://www.cs.toronto.edu/~tang/papers/mr_dbn.pdf. Basically, the variances are determined via approximate maximum likelihood learning, exactly like the other parameters (weight, biases) are learned.

(Mar 10 '13 at 08:15) osdf

Thanks. I scaled the data to have zero-mean and unit-variance. And then trained a GRBM, followed by multiple layers of binary RBMs. I am not sure what I am doing wrong, but all values after the first layer comes as NaN. My GRBM implementation is here.

(Mar 11 '13 at 04:29) Niharjyoti Sarangi

@osdf The problem got fixed after I reduced the learning rate by an order of magnitude.

(Mar 12 '13 at 02:04) Niharjyoti Sarangi

@niharjyoti, it's an old post, but do you remember the range in which your GBRBM reconstruction error fell in? I'm trying to train after normalizing my data to unit variance, zero mean, but my error always stays above around 0.60...

(Jul 11 '14 at 01:58) Muneeb
showing 5 of 6 show all
Be the first one to answer this question!
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.