The Fixed-Point Implementations for Recurrent Neural Networks

In this paper, the learning efficiency of the single-layer and multiple-layer locally recurrent neural networks (RNN) were investigated. In the RNN structure, piecewise linear activation functions were used. In addition, infinite impulse response digital filter played the role of signal recursions. In RNN implementation, pole-L 2 sensitivity minimization was performed. The weight of every neuron was adjusted by using the back-propagation (BP) learning algorithm. Simulation results show that multilayer RNN may have better learning performance. In addition, the RNN with optimal IIR filter implementation may have better learning performance than that of the canonical-form realization.