Balancing computational complexity and generalization ability: A novel design for ELM

Learning paradigms that use random basis functions provide effective tools to deal with large datasets, as they combine efficient training algorithms with remarkable generalization performances. The paper first considers the affinity between the paradigm of learning with similarity functions and the Extreme Learning Machine (ELM) model, and reformulates the mapping scheme of ELMs. A mapping scheme that better balances generalization ability and network size is a novelty point of the proposed approach, and represent a major advantage when targeting implementation on resource-constrained devices. A computationally efficient heuristic supports the training procedure, and suitably applies the theory of learning with similarity functions to the availability of consistent amounts of data. Experimental results on standard datasets confirm the effectiveness of the proposed approach.