I’m sorry if this question seems out of place, but I have tried asking this in other channels, and still wasn’t able to understand the PCD implementation in the tutorial. (The tutorial I’m referring to is https://github.com/lisa-lab/DeepLearningTutorials/blob/master/code/rbm.py) so I would appreciate some help here.
If I understand the logic correctly, PCD is using the last Gibbs sampled hidden layer of the previous batch/epoch(i.e. stored in ‘persistent’) to generate a reconstruction, and compare it to the input/visible layer of the current batch/epoch. How does that train the model if it’s comparing data from different batches? Or, is it that the two layers not being compared, but are completely independent of each other? (So the current batch is used to compute the positive phase of model to the hidden features, while the previous batch is used to generate the ‘imagination’ of the model at a random state?) -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to theano-users+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
