Recently, the notion that the brain is fundamentally a prediction machine has gained traction within the cognitive science community. Consequently, the ability to learn accurate predictors from experience is crucial to creating intelligent robots. However, in order to make accurate predictions 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. Often domain knowledge is employed to design useful features for specific problems, but learning representations in a domain independent manner is highly desirable. While many approaches for automatic feature extraction exist, they are often either computationally expensive or of marginal utility. On the other hand, methods such as Extreme Learning Machines (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 learning is done fully online, such as is the case for robots learning via their interactions with the world. Selectionist methods, 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 replication can improve performance beyond selection alone, and may offer a path towards effective learning of predictive models in robotic agents.