The Hamiltonian approach to neural networks dynamics
Ulrich Ramacher, P. Nachbar · 1991
The Hamiltonian concept of partial differential equation (PDE) theory is used to describe the dynamics of arbitrary neural nets. Pattern as well as learning dynamics are admitted simultaneously. Any method to determine the minima of the Hamiltonian with respect to the weight functions is shown to create an associated learning rule. Since arbitrary topologies and learning functions can be inscribed into the Hamiltonian, the concept turns out to be useful for a unified treatment of the dynamics of neural nets. Former results on the dynamics of special networks and learning functions are recovered.>