Feedback Error Learning with a noisy teacher

The feedback error learning (FEL) algorithm is examined under the condition of a noisy teaching signal. The teaching signal, which adaptively adjusts the weights of the feedforward network, is assumed to be corrupted by a signal dependent noise (SDN) source. The FEL framework was originally inspired by the cerebellum as a model for human motor control. We analyze the robustness properties of the original system with respect to the SDN noise model. We prove bounds on the learning rate and feedback gain matrices that guarantee stochastic stability of the closed loop system.