Online Extreme Evolutionary Learning Machines

Abstract

Recently, the notion that the brain is fundamentally a pre-diction machine has gained traction within the cognitive sci-ence community. Consequently, the ability to learn accu-rate predictors from experience is crucial to creating intel-ligent robots. However, in order to make accurate predic-tions it is necessary to find appropriate data representations from which to learn. Finding such data representations or features is a fundamental challenge for machine learning. Of-ten domain knowledge is employed to design useful features for specific problems, but learning representations in a do-main independent manner is highly desirable. While many approaches for automatic feature extraction exist, they are of-ten either computationally expensive or of marginal utility. On the other hand, methods such as Extreme Learning Ma-chines (ELMs) have recently gained popularity as efficient and accurate model learners by employing large collections of fixed, random features. The computational efficiency of these approaches becomes particularly relevant when learn-ing is done fully online, such as is the case for robots learn-ing via their interactions with the world. Selectionist meth-ods, which replace features offering low utility with random replacements, have been shown to produce efficient feature learning in one class of ELM. In this paper we demonstrate that a Darwinian neurodynamic approach of feature replica-tion can improve performance beyond selection alone, and may offer a path towards effective learning of predictive mod-els in robotic agents.