Extreme Learning Machine Design for Dealing with Unrepresentative Features

Extreme Learning Machines (ELMs) have become a popular tool for the classification of electroencephalography (EEG) signals for Brain Computer Interfaces. This is so mainly due to their very high training speed and generalization capabilities. Another important advantage is that they have only one hyperparameter that must be calibrated: the number of hidden nodes. While most traditional approaches dictate that this parameter should be chosen smaller than the number of available training examples, in this article we argue that, in the case of problems in which the data contain unrepresentative features, such as in EEG classification problems, it is beneficial to choose a much larger number of hidden nodes. We characterize this phenomenon, explain why this happens and exhibit several concrete examples to illustrate how ELMs behave. Furthermore, as searching for the optimal number of hidden nodes could be time consuming in enlarged ELMs, we propose a new training scheme, including a novel pruning method. This scheme provides an efficient way of finding the optimal number of nodes, making ELMs more suitable for dealing with real time EEG classification problems. Experimental results using synthetic data and real EEG data show a major improvement in the training time with respect to most traditional and state of the art ELM approaches, without jeopardising classification performance and resulting in more compact networks.