Restricted Bolzmann Machine for feature extraction

Question

I try to extact feature from Audio Signal PSD via a two-layer RBM. Then use the Feature as the input of SVM.

But what confuse me is that the output of RBM can be caculated by this formula : Y = X*W ? in which X is the input of

RBM,and Y is the output of RBM, W is the weight between the input(visual units) and output(hidden units).

Maybe the qustion is very easy for you, however I need the help, thank you.

Answer 1

anybody who can tell me : Why two RBMs did not work better than one RBM for the feature extract?

Answer 2

I have some progress now, for one two-layer RBM , the output should be sigm( C+ X * W), in which C is the bias term. sigm

is activation function.

But I have one more question: if I stack two RBMs together. What is the out put fomular? I have two idea, Please help to tell me which one is exact?

->1 Y1 = sigm( C1+ X * W1); Y2 = sigm( C2+ Y1 * W2);

->2 Y= sigm( C2 + X * W1 * W2);

    in which W1 is the 1st RBM weight and W2 is the 2nd RBM weight.

For My data, the 2nd fomular had better accuracy.
so who can do me a favour? thank you ~