Extreme learning machine (ELM) achieves faster training speed and higher classification accuracy, compared with other widely used classifiers, such as back propagation (BP), support vector machine (SVM), spectral clustering (SC), and so forth. However, ELM suffers from some drawbacks:1) ELM utilizes the calculation of inverse matrix for training, which cannot be adopted in the embedded system; 2) the training time of ELM increases dramatically for large-scale applications. To solve these drawbacks of ELM, a new training strategy called Sequential ELM (SELM) was proposed, which avoids the calculation of inverse matrix. Therefore, SELM can be adopted in the embedded system. It is proven that SELM achieves lower complexity than other widely used algorithms. Furthermore, simulations based on practical datasets indicate that the classification accuracy of SELM is higher than traditional ELM and other widely used classifiers with shorter training time.