What is an intuitive explanation of what a Hopfield network is?

However, the idea of building a generative model via an energy function is a good one. Since it is a probabilistic model, training a Boltzmann Machine often involves maximizing its likelihood function, which is proportional to the energy divided by the total energy of all configurations. This total energy is formally called the partition function, and is generally intractable. By restricting the connections of a Boltzmann Machine to be bipartite, people came up with Restricted Boltzmann Machine (RBM), whose inference can be approximated by MCMC and similar methods. Using RBMs to pre-train a deep feedforward nets is one of the main breakthroughs in Deep Learning several years ago.