Backpropagation Learning with Arctangent Activation Function

J. Kamruzzaman (Australia)

Keywords

Neural network, backpropagation learning, activation function, learning speed.

Abstract

The slow convergence of multiplayer feedforward network is partially due to the activation function used in learning algorithm. The commonly used activation functions have diminishing value of the first derivative as the nodes approach saturated values (0 or ± 1). In this paper, a new activation function is proposed to accelerate Backpropagation learning. Simulation using this activation function shows improvement in learning speed compared with other commonly used functions. This function may also be used in other multilayer feedforward training algorithms.

Important Links:



Go Back


IASTED
Rotating Call For Paper Image