Two-stage extreme learning machine for high-dimensional data

Extreme learning machine (ELM) has been proposed for solving fast supervised learning problems by applying random computational nodes in the hidden layer. Similar to support vector machine, ELM cannot handle high-dimensional data effectively. Its generalization performance tends to become bad when it deals with high-dimensional data. In order to exploit high-dimensional data effectively, a two-stage extreme learning machine model is established. In the first stage, we incorporate ELM into the spectral regression algorithm to implement dimensionality reduction of high-dimensional data and compute the output weights. In the second stage, the decision function of standard ELM model is computed based on the low-dimensional data and the obtained output weights. This is due to the fact that two stages are all based on ELM. Thus, output weights in the second stage can be approximately replaced by those in the first stage. Consequently, the proposed method can be applicable to high-dimensional data at a fast learning speed. Experimental results show that the proposed two-stage ELM scheme tends to have better scalability and achieves outstanding generalization performance at a faster learning speed than ELM.