Extending Extreme Learning Machine with Combination Layer

We consider the Extreme Learning Machine model for accurate regression estimation and the related problem of selecting the appropriate number of neurons for the model. Selection strategies that choose “the best” model from a set of candidate network structures neglect the issues of model selection uncertainty. To alleviate the problem, we propose to remove this selection phase with a combination layer that takes into account all considered models. The proposed method in this paper is the Extreme Learning Machine(Jackknife Model Averaging), where Jackknife Model Averaging is a combination method based on leave-one-out residuals of linear models. The combination approach is shown to have better predictive performance on several real-world data sets.