(22th-December-2020)
• Similarly to Boltzmann machines, denoisingautoencoders and their generalizations (such as GSNs, described below) can be used to sample from a conditional distribution p(xf | xo), simply by clamping the observed units xf and only resampling the free units xo given xf and the sampled latent variables (if any). For example, MP-DBMs can be interpreted as a form of denoisingautoencoder, and are able to sample missing inputs. GSNs later generalized some of the ideas present in MP-DBMs to perform the same operation ( , ). ( ) Bengio et al. 2014 Alain et al. 2015 identified a missing condition from Proposition 1 of ( ), which is Bengio et al. 2014 that the transition operator (defined by the stochastic mapping going from one state of the chain to the next) should satisfy a property called detailed balance, which specifies that a Markov Chain at equilibrium will remain in equilibrium whether the transition operator is run in forward or reverse.
Comments