Class of Hamiltonian neural networks
Philippe De Wilde · Physical review. E, Statistical physics, plasmas, fluids, and related interdisciplinary topics · 1993
We investigate analog neural networks. They have continuous state variables that depend continuously on time. Although they all have an energy function, not all can have their dynamics derived from a Hamiltonian. Some necessary conditions are given for the network to have Hamiltonian dynamics. We give an example and, using symplectic transformations, describe a whole class of neural networks with Hamiltonian dynamics.