I am training Restricted Boltzmann MAchine with MNIST Datasets. As per my understanding if it can reconstruct input proper then i will say it learnt. But, when i use Persistent Contrastive Divergence, it will not reconstruct proper because positive and negative phase samples doesnot match. In this case, how i will find, it learnt or not?
Even when use Persistent Contrastive Divergence, the RBM will be able to reconstruct its input. If it is not the case, you may have a problem in your implementation of PCD.
You should know that the reconstruction error rate is not a very good proxy of Log-Likelihood for PCD. If you use it to monitor the progress of the training, you should consider using the pseudo likelihood.