The idea of Hinton et al's algorithm

• Train the autoencoder. That's the two layer Boltzmann machine.
• That should now self-correct.
• Repeatedly add new layers to improve performance.
• Hinton then refines the weights to improve results.