DBNs are a class of unsupervised probabilistic/graphical deep learning algorithms. The goal of a DBN is to classify data into different categories. They are composed of multiple layers of stochastic latent variables, which can be referred to as feature detectors or hidden units. It is these hidden units that capture correlations present in the data.
DBNs were introduced in 2006 by Geoffrey Hinton and have since been widely used in the following areas:
- Image recognition, generation, and clustering
- Speech recognition
- Video sequences
- Motion capture data
Before trying to fully understand a DBN, there are two fundamental notions to be considered and understood:
- Bayesian Belief Networks (BBNs)
- Restricted Boltzmann machines (RBMs)
