We consider the problem of learning to estimate depth from stereo image pairs. This can be formulated as unsupervised learning - the training pairs are not labeled with depth. We have formulated an algorithm which maximizes conditional likelihood the left image given right image in a model that involves latent information (depth). This unsupervised learning algorithm implicitly trains shape from texture and shape from shading monocular depth cues. The talk will present pragmatic results in the stereo vision problem as well as a general formulation of models and methods for maximizing conditional likelihood in a latent variable model where we wish to interpret the latent information as "labels".