1: \begin{abstract} 

2:   Restricted Boltzmann Machines (RBMs) are general unsupervised

3:   learning devices to ascertain generative models of data

4:   distributions. RBMs are often trained using the Contrastive

5:   Divergence learning algorithm (CD), an approximation to the gradient

6:   of the data log-likelihood. A simple reconstruction error is often

7:   used as a stopping criterion for CD, although several

8:   authors~\cite{schulz-et-al-Convergence-Contrastive-Divergence-2010-NIPSw,

9:     fischer-igel-Divergence-Contrastive-Divergence-2010-ICANN} have

10:   raised doubts concerning the feasibility of this procedure.  In many

11:   cases the evolution curve of the reconstruction error is monotonic

12:   while the log-likelihood is not, thus indicating that the former is

13:   not a good estimator of the optimal stopping point for learning.

14:   However, not many alternatives to the reconstruction error have been

15:   discussed in the literature. In this manuscript we investigate

16:   simple alternatives to the reconstruction error, based on the

17:   inclusion of information contained in neighboring states to the

18:   training set, as a stopping criterion for CD learning.

19: \end{abstract}