Notes Paper decreibes using a Gaussian distruibuted visible layer. Code (and reality) says this should be a logistic unit. The learnign weights <v.h> should use probabilites of h rather than their binary sampled values.