Composition methods for the integration of dynamical neural networks

摘要

We apply the symmetric composition method for the integration of ordinary differential equations to dynamical neural networks. In this method, we split the vector field, which is parameterized by a neural network, into the contribution of each of its neurons. We then solve the elementary differential equation associated to each neuron separately, and recombine these contributions in a sequence of compositions. This gives rise to simple integration rules for dynamical neural networks, which we present for dynamical single-hidden-layer perceptrons.