Target adaptive extreme learning machine for transfer learning

Extreme learning machines (ELM) have been applied in several fields due to their simplicity and computational efficiency. However, ELM hurts the performance in cross-domain learning problems similar to most machine learning algorithms. In this paper, we mainly focus on the semi-supervised transfer learning algorithm under ELM framework. Unlike other transfer learning methods employed both source and target domains, we propose a target adaptive ELM (TAELM) of learning a high-quality target-specific classifier with less resources. We formulate a novel objective function to obtain a target-specific classifier by introducing a knowledge transfer term on a pre-trained source model and a graph laplacian-based manifold regularization term on the target domain, while its solution are analytically determined without loss of the computing efficiency and learning ability of traditional ELM. In our experiments, we verify the effectiveness of the proposed approach by using a deep neural network model as feature extractor for both domains. Experimental results demonstrate that our method with less resources significantly outperforms other state-of-the-art algorithms.