What does a neural network named after a physicist who worked on the atomic bomb actually do?

The answer

Boltzmann Machines, introduced by Geoffrey Hinton and Terrence Sejnowski in 1985, are stochastic recurrent neural networks that learn internal representations through a process inspired by statistical mechanics. Named after Ludwig Boltzmann and his work on the probability distribution of particle energies, these networks use simulated annealing - gradually reducing random noise - to find optimal configurations. Each neuron makes probabilistic decisions based on the weighted sum of its inputs and a temperature parameter, allowing the network to escape local minima and explore the solution space. Though full Boltzmann Machines proved computationally impractical due to extremely slow learning, the Restricted Boltzmann Machine variant later became a key building block for deep learning, particularly in Hinton's 2006 deep belief networks.