Restricted Boltzman Machines question

[sergio.rodri...@gmail.com]

Hi Cesar,

I have a question regarding the RBMs implemented in the last version of Accord.NET: Is it possible to use "double" as inputs of a Deep NN? In your example of DNN, the input is a binary image.

I also read in some papers that the learning algorithm for binary or real-valued hidden layers (of RBM) is different, but I don't know if it is the same for the input of the network. I am not sure if the implementation of the Contrastive Divergence take this into account.

By the way: Thanks for the implementation!

[César]

Hi there!

Thanks for the interest in the framework!

It is possible to pass double[] as inputs for a deep neural network, though most likely you will need to specify a Gaussian activation function for the network (by the way, sorry for the not-so-useful documentation page. I am going to amend this in the next release).

To specify a Gaussian function, you can either use one of the network constructors which accept an IStochasticFunction and pass a Gaussian function, or use the CreateGaussianBernoulli static method for a shorter way:

// Creates a Gaussian-Bernoulli network with 5 inputs, 10 hidden neurons
// in the first hidden layer, and 1 output neuron in the last hidden layer
DeepBeliefNetwork ann = DeepBeliefNetwork.CreateGaussianBernoulli(5, 10, 1);

Hope it helps!

Best regards,

Cesar

[sergio.rodri...@gmail.com]

Thanks! As far as I understande, I will need to normalize the data in order to use the Gaussian Function, won't I? Or does the Gaussian Function normalize the data inside?

[César]

Yeah, the data needs to be normalized first to have zero mean and unit variance. By the way, please keep in mind the Gaussian feature is still a bit experimental; though I and others already had some success with it, it is still a relatively new feature (which might need more user testing).

[sergio.rodri...@gmail.com]

ok thanks!