A modified batch intrinsic plasticity method for pre-training the random coefficients of extreme learning machines

In extreme learning machines (ELM) the hidden-layer coefficients are randomly set and fixed, while output-layer of neural network computed by a least squares method. The randomly-assigned in ELM known to influence its performance accuracy significantly. this paper we present modified batch intrinsic plasticity (modBIP) method for pre-training random networks. current is devised based on same principle as (BIP) method, namely, enhancing information transmission every node network. It differs from BIP two prominent aspects. First, modBIP does not involve activation function algorithm, it can be applied with any contrast, employs inverse construction, requires invertible (or monotonic). work often-used non-monotonic functions (e.g. Gaussian, swish, Gaussian error linear unit, radial-basis type functions), which breaks down. Second, generates target samples intervals minimum size, leads highly accurate computation results when combined ELM. ELM/modBIP markedly more than ELM/BIP numerical simulations. Ample experiments presented shallow deep networks approximation boundary/initial value problems partial differential equations. They demonstrate that produces simulation results, insensitive random-coefficient initializations This sharp contrast without coefficients.