In PCA and LDA, we had severe limits to the number of components we were allowed to extract. For PCA, we were capped by the number of original features (we could only use less than or equal to the number of original columns), while LDA enforced the much stricter imposition that caps the number of extracted features to the number of categories in the ground truth minus one.

The only restriction on the number of features RBMs are allowed to learn is that they are limited by the computation power of the computer running the network and human interpretation. RBMs can learn fewer or more features than we originally began with. The exact number of features to learn is up to the problem and can be gridsearched.