Complex probabilistic models of unlabeled data can be created by combining simpler models. Mixture models are obtained by averaging the densities of simpler models and "products of experts" are obtained by multiplying the densities together and renormalizing. A far more powerful type of combination is to form a "composition of experts" by treating the values of the latent variables of one model as the data for learning the next model. The first half of the tutorial will show how deep belief nets -- directed generative models with many layers of hidden variables -- can be learned one layer at a time by composing simple, undirected, product of expert models that only have one hidden layer. It will also explain why composing directed models does not work. Deep belief nets are trained as generative models on large, unlabeled datasets, but once multiple layers of features have been created by unsupervised learning, they can be fine-tuned to give excellent discrimination on small, labeled datasets. The second half of the tutorial will describe applications of deep belief nets to several tasks including object recognition, non-linear dimensionality reduction, document retrieval, and the interpretation of medical images. It will also show how the learning procedure for deep belief nets can be extended to high-dimensional time series and hierarchies of Conditional Random Fields.