Mike Hasselmo and colleagues examined how the brain generalizes and infers new behaviors from previous experience. They trained different styles of neural network models to learn context-dependent behaviors (i.e., the response to four stimuli, A B C D, mapped onto two different responses X Y differently in different contexts). There were previously unseen stimuli whose response could be inferred from the other stimuli. They analyzed a Deep Belief Network, a Multi-Layer Perceptron, and the combination of a Deep Belief Network with a Linear Perceptron. The combination of the Deep Belief Network with Linear Perceptron worked best.
A Deep Belief Network has multiple layers of hidden units with connections between, but not within, the layers.