Extreme Learning Machine with Adaptive Growth of Hidden Nodes and Incremental Updating of Output Weights

The extreme learning machines (ELMs) have been proposed for generalized single-hidden-layer feedforward networks (SLFNs) which need not be neuron alike and perform well in both regression and classification applications. An active topic in ELMs is how to automatically determine network architectures for given applications. In this paper, we propose an extreme learning machine with adaptive growth of hidden nodes and incremental updating of output weights by an error-minimization-based method (AIE-ELM). AIE-ELM grows the randomly generated hidden nodes in an adaptive way in the sense that the existing hidden nodes may be replaced by some newly generated hidden nodes with better performance rather than always keeping those existing ones in other incremental ELMs. The output weights are updated incrementally in the same way of error minimized ELM (EM-ELM). Simulation results demonstrate and verify that our new approach can achieve a more compact network architecture than EM-ELM with better generalization performance.