Transfer Extreme Learning Machine with Cross Domain Mean Approximation Projection

Abstract

Traditional extreme learning machine (ELM) with the advantages of fast speed and high precision, is widely used in pattern recognition and machine learning. However, it has two shortcomings: 1) Its accuracy decreases greatly under lack of available training samples. 2) The performance is greatly affected by the initialization of hidden-layer parameters. To solve the above two problems, we propose a novelty ELM named transfer extreme learning machine with cross domain mean approximation projection (TELM-CDMAP). It first uses cross domain mean approximation (CDMA) to project input data into feature subspace instead of the data transformation from input-layer to hidden-layer. Then, we apply a between-domain scatter matrix (BDSM) term into the hidden-layer and solve the optimal output weight with the ability of knowledge transfer. The classification experiments on Office+Caltech object recognition and USPS+MNIST digital handwriting datasets are carried out and the result show the effectiveness of our ELM models.