DBNs are generative models, and usually you sample by thermalising the deepest layer (as it's a restricted Boltzmann Machine), and then forward propagating a sample towards the visible layer to get a sample from the learned distribution.

This is less flexible sampling than in a single layer DBN: a restricted Boltzmann Machine. There, we can start our sampling chain at any state we want, and get samples "around" that state. In particular, we can clamp some visible nodes $\{v_i\}$ and get samples from the conditional probability $p(v_j|\{v_i\})$.

Is there a way to do something similar in DBNs? When we interpret the non-RBM layers as RBMs by removing directionality, can we treat it as a Deep Boltzmann Machine and start sampling at e.g. a training example again?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.