Commonly NN (autoencoders) use a set of weights in the reduction process and another in the reconstruction process.

But a RBM uses the same weights in construction and reconstruction process. I don't think that it can create inverse function of construction using the same weights.

Why can Restricted Boltzmann Machines reconstruct using the same weights?


Restricted Boltzmann Machines are particular generative models. When you define the model with some parameters (weights, bias...) this means that you put in an energy function and you define the joint law of the visible and hidden layer $p(\mathbf{v},\mathbf{h})$. This is the first step.

Then, this law contains enough information to deduce $p(\mathbf{h}|\mathbf{v})$ and $p(\mathbf{v}|\mathbf{h})$ which are used in the reconstruction process.

A nice complete document on the topic is Training restricted Boltzmann machines, by Asja Fischer which can be found here.

| cite | improve this answer | |
  • $\begingroup$ thanks for telling doc. i don't know completely, but maybe found it was relevant to static state of output value in hopfield network. $\endgroup$ – cartman May 13 '19 at 10:06
  • $\begingroup$ ok, I maybe found to make static state of two distributions using same weights than calculating input v, hidden h value. $\endgroup$ – cartman May 13 '19 at 10:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.