C-ELM: A Curious Extreme Learning Machine for Classification Problems

Abstract

In psychology, curiosity is generally known as the critical intrinsic motivation for learning. It drives human beings to explore for novel and interesting information that can elicit the feeling of pleasure. This paper proposes such a curiosity driven algorithm for Extreme Learning Machine, which is referred to as Curious Extreme Learning Machine (C-ELM). C-ELM follows the psychological theory of curiosity proposed by Berlyne and performs curiosity appraisal towards each input data based on four collative variables: novelty, uncertainty, conflict, and surprise. The collative variables reflect the level of curiosity stimulation in the input data. Based on the level of curiosity stimulation, the network decides on the strategies for learning, including neuron addition, neuron deletion, and parameter update. During neuron addition, a new neuron is added based on the input data, thereby reducing the randomization effect of ELM. The parameter update is conducted using recursive least squares method and neuron deletion aims at deleting the most conflicting knowledge. The empirical performance study of the proposed method on benchmark classification problems clearly highlights the learning and generalization ability of C-ELM.