Adjoint-operators and non-adiabatic learning algorithms in neural networks (Q804485)

Various schemes for temporal learning have been proposed. Their drawbacks usually comprise repeated solutions of (weights, gains) gradients, storing of a large amount of history data or adiabatic assumptions. This paper proposes a new rigorous methodology enabling the implementation in a highly efficient manner. It combines the advantage of dramatic complexity reductions inherent in adjoint methods with the ability to solve the equations forward in time without adiabatic assumptions, so that real-time applications are possible. All involved mathematics is described, while no particular application is demonstrated.