A robust learning algorithm for feedforward neural networks with adaptive spline activation function

This paper proposed an adaptive robust learning algorithm for spline-based neural network. Adaptive influence function was dynamically added before objective function to modify the learning gain of back-propagate learning method in neural networks with spline activation functions. Besides the nonlinear activation functions in neurons and linear interconnections between neurons, objective function also changes the shape during iteration. This employed neural network the robust ability to reject gross errors and to learn the underlying input-output mapping from training data. Simulation results also conformed that compared to common learning method, convergence rate of this algorithm is improved for: 1) more free parameters are updated simultaneously in each iteration; 2) the influence of incorrect samples is gracefully suppressed.