LAUSR

Multimodal data have recently become nearly ubiquitous in the real world. Exploring the multimodal fusion is beneficial to improve the performance of the system. However, it is difficult to ensure that data collected from different sources are full pairing. In this paper, we will focus on the weakly paired case of multimodal data, i.e., each modality is partitioned into multiple groups, only paired information on groups is known instead of full pairing between data samples. A new framework of weakly paired multimodal fusion based on multilayer extreme learning machine (ML-ELM) is proposed in this paper, which will find complex nonlinear transformations of each modality of data such that the resulting representations are highly correlated. In this framework, unsupervised hierarchical ELM performs feature extraction for all modalities separately. Then, the higher-level representations from all modalities perform joint dimension reduction by weakly paired maximum covariance analysis. We evaluate our framework on three challenging cross-modal datasets, and the results have proved the effectiveness of proposed method.