Online Sequential Extreme Learning Machine based Instinct Plasticity for Classification

Abstract

Random determination of input weights leads to unstable performance in Online Sequential Extreme Learning Machines (OS-ELM), so obtaining reliable input weights was expected to improve the model performance. We designed a new model—the OS-ELM based Instinct Plasticity with a new weight selection scheme (NOS-ELM-IP) to enhance the forecast stability and accuracy for classification. In this model, the input weights were selected by a new weight selection method, which replaced the original random selection part in OS-ELM. Moreover, the Instinct Plasticity idea was used to find the gain and bias, used in the sequential training part of OS-ELM. It maximized the information of hidden neurons and enlarged the memory. The experimental results show that the proposed new weight selection method and Instinct Plasticity rule enhanced the overall performance in classification tasks for binary and multi-class data sets.