基于超限学习机的非线性典型相关分析及应用

典型相关分析是目前常用的研究两个变量间相关性的统计方法。针对线性典型相关分析难以准确揭示变量之间复杂关系的问题，提出一种基于超限学习机的非线性典型相关分析多模态特征提取方法。首先，采用超限学习机分别的对每个模态进行无监督特征学习，得到抽象的深度特征表示；然后将这些深度抽象特征通过典型相关分析极大化模态之间的相关性，同时得到两组相关变量，实现多模态数据的复杂非线性和高相关性表示。最后在康奈尔大学机器抓取公开数据集上进行实验验证，结果表明，所提出的方法与其他相关算法相比，训练速度得到显著提升。

Nonlinear canonical correlation analysis and application based on extreme learning machine

Canonical correlation analysis (CCA) is a statistical technique commonly used to determine the correlativity of two variables. It is difficult to accurately identify the complex underlying relationship between variables using linear CCA, so we propose a nonlinear CCA based on an extreme learning machine (ELM) for multi-modal feature extraction. First, to obtain abstract-depth feature representation, we use the ELM to perform unsupervised feature learning for each modality. Then, we use CCA to maximize the correlation between the nonlinear representations, thereby simultaneously obtaining two groups of related variables, and realize complex nonlinear and high-correlation representations of multi-modality data. Lastly, we conducted an experiment using the Cornell grasping dataset. The results show that, in comparison with other related algorithms, the proposed method significantly increases the training speed.