Predicting (function approximation) is perhaps the main use of neural nets but there is another very important use: generating, or in other words, imagination. And in generative mode we need to know about probabilities and latent variables - thus, we need a little more than neural nets to do this task.
