X behaves like a prior. Most guesses are good guesses if there is no prior knowledge of what X should be. We should view the algorithm as a non-supervising algorithm similar to clustering. Messages are passing through the connection so the latent feature will reach a stable state.