Extreme learning machine versus classical feedforward network

Abstract Our research is devoted to answering whether randomisation-based learning can be fully competitive with the classical feedforward neural networks trained using backpropagation algorithm for classification and regression tasks. We chose extreme as an example of networks. The models were evaluated in reference training time achieved efficiency. conducted extensive comparison these two methods various tasks scenarios: $$\bullet$$ ∙ comparable network capacity architectures tuned each model. was on multiple datasets from public repositories some artificial created this research. Overall, experiments covered more than 50 datasets. Suitable statistical tests supported results. They confirm that relatively small datasets, machines (ELM) are better by algorithm. But demanding image like ImageNet, ELM not modern backpropagation; therefore, order properly address current practical needs pattern recognition entirely, further development. Based our experience, we postulate develop smart algorithms inverse matrix calculation, so determining weights challenging becomes feasible memory efficient. There a need create specific mechanisms avoid keeping whole dataset compute weights. These most problematic elements processing, establishing main obstacle widespread application.