Tensor Decomposition Based Approach for Training Extreme Learning Machines

Conventional Extreme Learning Machines utilize Moore–Penrose generalized pseudo-inverse to solve hidden layer activation matrix and perform analytical determination of output weights. Scalability is the major concern to be addressed in Extreme Learning Machines while dealing with large dataset. Motivated by these scalability concerns, this paper proposes a novel tensor decomposition based Extreme Learning Machine which utilize PARAFAC and TUCKER decomposition based techniques in a SPARK platform. This proposed Extreme Learning Machine achieve reduced training time and better accuracy when compared with a conventional Extreme Learning Machine.