Harnessing chaotic activation functions in training neural network

Abstract

We propose'harnessed Chaotic Activation Functions' (HCAF) to compute final activation of a neural network. That is biologically plausible to connect with neuron. Multilayer feed-forward neural networks are trained with a supervised algorithm which is loosely connected with biological learning. Bio-inspired system development is recently a challenging topic in intelligent system design. We investigate whether HCAF can enable the learning to be faster. Validity of the proposed method is examined by performing simulations on challenging five real benchmark classification problems. The HCAF has been examined to 2-bit, Diabetes, Wine, Glass and Soybean problems. The algorithm is shown to work better than other AFs used independently in BP such as sigmoid(SIG), arctangent (ATAN), logarithmic (LOG), robust chaos in neural network (RCNN), and that of jointly such as fusion of activation function (FAF). © 2012 Springer-Verlag.