A Joint Unsupervised Cross-Domain Model via Scalable Discriminative Extreme Learning Machine

Abstract

Extreme learning machine (ELM) is a well-known cognitive model, that has been extended to cross-domain tasks. Nonetheless, most existing paradigms that are based on ELM either concern about a case in which a specific number of instances are labelled in the target domain or a learner is trained without sufficient capacity to eliminate the gap between domains. To cope with the scenario in which there are no target labels and to acquire a better adaptive learner, we propose a joint unsupervised cross-domain model via scalable discriminative ELM, which is abbreviated as JUC-SDELM. Within the framework, the scalable factor is integrated into discriminative ELM (DELM) to adjust the output margin, which strengthens the discriminative capacity of the ELM classifier. In addition, we follow the basic strategy of joint distribution adaptation (JDA) to align the subspaces generated by JUC-SDELM in terms of their statistics. The discrepancy across domains is alleviated after a few iterations. Moreover, a metric on the outputs of ELM is utilized to filter unreliable pseudo labels in the target domain, with the aim of eliminating the negative transfer effect. Results are obtained by comparing JUC-SDELM with state-of-the-art baseline methods on 16 cross-domain benchmarks that were constructed based on three combined datasets. Likewise, the outcomes in terms of key parameters are also examined. According to the experiments, our proposed model achieves competitive overall performance.