Compact Extreme Learning Machines for biological systems

In biological system modelling using data-driven black-box methods, it is essential to effectively and efficiently produce a parsimonious model to represent the system behaviour. The Extreme Learning Machine (ELM) is a recent development in fast learning paradigms. However, the derived model is not necessarily sparse. In this paper, an improved ELM is investigated, aiming to obtain a more compact model without significantly increasing the overall computational complexity. This is achieved by associating each model term to a regularized parameter, thus insignificant ones are automatically unselected, leading to improved model sparsity. Experimental results on biochemical data confirm its effectiveness.