Implementation of activation functions for ELM based classifiers

Abstract

Artificial neural network is an information processing system which is inspired by biological neural system. It is widely utilized because of its competency to derive consequential information from complex data, adaptive learning, self organization and real time operation. One of the most promising task of neural network is classification and prediction. Extreme learning machines (ELM), a single layer feed forward network is used for this purpose because of its advantage over conventional neural network; i.e. better generalization performance, ability to evade the presence of local minima and universal approximation theorem. Depending on the types of activation functions used in ELM structure, the performance of the network varies. Implementation of activation functions in FPGA is a challenging task as it has to be digitized in order to represent it in hardware. A comparison study is performed predicated on the hardware requisite and speed of the architecture for different activation functions. Withal, a comparison study is given based on the training and testing precision of each activation function.