|
【book】【Ahmed Azab】A review on transfer learning approaches in brain–computer interface
可以用的句子
需要长期校准的原因
每个人大脑模式不同
非平稳
高噪声
迁移的原则
迁移学习的类型
归纳迁移学习
The purpose of inductive transfer learning algorithms is to improve estimation of thetarget predictive function
fT(.) in target domain when the target and source tasks aredifferent (TD~= TS).
直推式迁移学习
The goal of transductive transfer learning algorithms is to improve estimation of thetarget predictive function
fT(.) in the target domain when the target and source tasksare same, but the target and source domains are
different [10].It is noted that in thetransductive transfer learning we assume no or few labelled trials are available
in thetarget domain whereas a large amount of labelled trials are available in the sourcedomain.
未监督迁移学习
This type of transfer learning tries to solve the learning problem when there are nolabelled trials available in
both the source and target domains during training. In theunsupervised transfer learning, while both the source
and target tasks are different, there is a relation between them. Unsupervised transfer learning algorithms can
beapplied to problems involving clustering and dimensionality reduction [10].
迁移学习方法
实例迁移
重要性采样
基于协变量自适应袋装重要性加权LDA（袋装IWLDA）
主动学习算法
mSDAs
重新加权
KL散度
特征表示迁移
CSP特征表示
Non CSP-based feature-representation transfer learning
非负矩阵分子
PCA
基于分类器的迁移
域自适应
DASVM
using source domain to initialise the discriminative function;
replace samplesfrom source domain with samples from target domain to adjust the
discriminativefunction
仅使用来自目标域的数据学习最终辨别函数。
集成学习
使用受试者的CSP滤波器组为每个受试者学习两套稀疏的滤波器组，称为稳健滤波器组和自适应滤波器组。
两个分类模型训练基于这两个滤波器组每个主题;
采用两级集合策略，将来自稳健集合模型和适应性集合模型的所有分类器整合为一个稳健集合学习器和一个适应性集合
学习器。
基于关系的迁移
挑战和讨论
|