Characteristics of gradient descent learning with neuronal gain control

The human brain shows the capability of adjusting the gain of neurons at the sensory periphery. In this paper, we investigate the properties of the backpropagation learning algorithm with adaptive neuronal gain, and compare its performance with the conventional one, and with the one combining dynamic learning rate optimization. Simulation results have shown that the algorithm can achieve the goal of fast convergence, and can alleviate the problem of local minima with a moderate increment of computation and storage burden