Complexity control methods for recurrent neural networks

摘要

We demonstrate that a dynamical complexity of neural networks can be controlled by our gradient-based method. An essential core of our method is calculation of the complexity and its partial derivative of the network parameters. Simulation results show that our method can make the exponent converge to desired values. In addition, the networks can learn not only the target time series, but also the complexity of the target by a combination of proposed method and conventional learning algorithm.