基于高斯过程隐变量模型的高维小样本多模态分类问题研究

Classification plays a key role in the pattern recognition. Compared with single-view based methods, multi-modal learning often obtains better performance in the classification, due to the exploitation of the correlation among different modalities. However, (1) most of existing methods need to estimate a large number of parameters, resulting in over-fitting and classification performance degradation if the multi-modal data is high-dimensional and small-sample-size; (2) they ignore the case that different modalities may miss partial samples; (3) Fews of them are able to represent the high-dimensional and small-sample-size multi-modal data deeply. To address these problems, the Gaussian Process Latent Variable Model (GPLVM) based multi-modal learning method is studied in this proposal. The main contribution is that the auto-encoder, deep learning, multi-kernel learning, laplacian prior, and large margin prior are jointly introduced into the model to obtain the Auto-Encoder Multi-Modal Gaussian Process Latent Variable Model. Our proposed method is capable of fully exploiting the supervised information, learning the classifier online, processing the incompleted multi-modal data, and deeply representing the input with only a few of parameters. Compared with existing multi-modal methods, it is more flexible and enjoys more powerful data classification capacity. Finally, we apply the proposed multi-modal method to biometric recognition and disease detection etc. to demonstrate its effectiveness.

分类是模式识别领域中的重要分支。相比于传统的单模态学习方法，多模态学习方法利用不同模态数据间的相关性信息，往往能在分类应用上获得更优的效果。然而现有算法（1）需估计大量的参数，在处理高维小样本多模态数据上极易过拟合，从而使分类器性能急剧下降；（2）忽略了部分样本对应的某一或某多种模态数据缺失的问题；（3）对高维小样本多模态数据的深度表示鲜有研究。为解决上述问题，本项目提出了基于高斯过程隐变量模型的多模态学习方法。其创新之处在于结合自编码、深度学习、多核学习、拉普拉斯先验和最大边缘先验，建立自编码多模态高斯过程隐变量模型。该模型可充分利用训练样本中的有监督信息和在线学习分类器，可处理样本缺失问题，同时可只用少量参数对高维数据进行深度表示。因此该多模态模型具有更强的灵活性和数据分类能力。最后，本项目所提模型将应用于生物特征识别和医学诊断等高维小样本多模态数据的实际分类中，以验证其高效性。

随着多媒体技术的不断发展，具体事物的表示形式也逐渐趋于多样化。通常这类多样化数据被称为多模态数据，而针对此数据提出的研究方法被称为多模态或多视角学习。相较于单模态或单视角学习，多模态学习算法充分利用了多模态数据中的相关性和互补性信息，在诸多任务中都得到了较大的性能提升。团队自承担该项国家自然科学基金项目以来，从多模态高维小样本数据为研究对象入手，从少量模型参数、模态缺失、深度学习等角度出发，提出了多种结合高斯过程隐变量模型、混合高斯模型、深度学习的多模态融合算法，并将其成功应用于目标识别、疾病诊察、缺失补全等任务中，并通过在多个数据集上的应用，证明了本项目组所提出的算法和模型在相应的任务上的有效性。..基于项目研究成果，共计出版英文专著1部，发表论文15篇，包括SCI检索国际期刊论文13篇，CCF A类国际会议论文2篇。申请发明专利4项，培养博士研究生3名，硕士研究生3名。建立了基于无创式体外诊察信息的多模态疾病诊察系统。上述指标均达到或超过项目预定目标。