When we try to use RBM as generative model to classify or analysis data. After long time of training, the model can learn the features of primary data very well, how could I select the significant features of the input by analysis the hidden units?
What I mean is, how could I use RBM model as a method to reduce the input dimension without converting those dimensions to other format(as Hinton, Geoffrey E. 2006), just select the several significant features/dimensions from the original dimensions?
http://www.cs.toronto.edu/~hinton/science.pdf
Refer Backpropagation neural networks technique for understand the concept of hidden and visible layers by "An introduction of ANN by S.N. Sivanandam"
I'm not sure that you really can do that with an RBM. Because the hidden layer is fully connected to the visible layer, the RBM learns a distributed representation of the input. That is, each hidden unit could draw from a whole bunch of the visible units, so we can't say that that hidden unit is really activated by only a single visible unit.
I think what you're interested in is feature selection (http://en.wikipedia.org/wiki/Feature_selection), where you try to learn which subset of your input features is most predictive.
Thanks mahendra and Zachary's answers.
what I work is want to verify/seek/fumble the meaning of the bunch of visible units for each hidden unit(I will pick up some hidden units for analysis). Do you think whether it is worthy for try, based on your research background?
As far as I‘m concerned, what I do is related with representation about hidden units of RBM. Do you have any articles related with it for recommending?
- Badges
- Science topic