模式识别 - 多视角学习 (Multi-View Learning) 简介

版权所有, 如需引用和转载, 请站内联系.

 

多视角学习 (Multi-View Learning) 简介

 

Views from (1) multiple sources (2) different feature subsets;

视角来自于 (1) 多个源 (2) 多个特征子集

 

Multi-view learning algorithms: (1) co-training (2) multi-kernel learning (3) subspace learning;

多视角学习算法: (1) 协同训练 (2) 多核学习 (3) 子空间学习

Co-training(协同训练) : 

trains alternately to maximizethe mutual agreement on two distinct views of theunlabeled data.

Multiple kernel learning(多核学习 ) : 

exploits kernels that naturallycorrespond to different views and combine kernels either linearly ornon-linearly to improve learning performance.

Subspace learning(子空间学习 ) : 

obtains a latent subspace shared bymultiple views by assuming that the input views are generated fromthis latent subspace.

 

Principles: (1) consensus principle (共识准则) (2) complementary principle (互补准则);

准则: (1) 共识准则 (2) 互补准则

 

Multi-view learning: introduces one function to model a particular view and jointly optimizes all the functions to exploit the redundant views of the same input data and improve the learning performance.

多视角学习:  引入了一个函数去模型化一个特定的视角, 并且利用相同输入的冗余视角去联合优化所有函数, 最终提高学习效果.

 

Co-training: trains alternately(轮流地) to maximize the mutual agreement on two distinct views of the unlabeled data.

协同训练: 在未标记数据的两个不同视角下, 轮流的训练, 使相互一致性最大化.

        Co-training Variants(变种):

                  (1) Expectation-Maximization (EM): assigning changeable probabilistic labels to unlabeled data; (聚类算法-OpenCV有相关函数)

                  (2) Semi-Supervised Learning Algorithm: Muslea et al;

                  (3) Bayesian undirected graphical model & Gaussian process classifiers: Yu et al;

                  (4) Combinative label propagation (结合的标签传播): Wang & Zhou;

                  (5) A data-dependent “co-regularization(协同正则化)”norm: Sindhwani;

                  (6) Data clustering and designed effective algorithms: Bickel & Scheffer and Kumar et al;

         Relying on three assumptions (依赖的三个假设):

                  (1) sufficiency (充足性) - each view is sufficient for classification on its own;

                  (2) compatibility (兼容性) – the target function of both views predict the same labels for co-occurring features with a high probability;

                  (3) conditional independent (条件独立性) – views are conditionally independent given the label.