脑-机接口系统的类协同式半监督学习

Like-co-training Semi-supervised Learning for Brain-computer Interface

针对如何提高有标签样本不足时的分类精度问题。提出脑-机接口系统(BCI)的类协同半监督学习算法(LCTSSL),采用有监督和无监督两种算法提取双特征训练双分类器协同扩充有标签样本集。在训练前后阶段设置不同置信度度量,选择两分类器分类结果一样的高置信度样本进行标记,保持每类每次新标记样本数目一样,提高有标样本集的可信度及识别系统的鲁棒性。迭代更新两分类器、有监督提取系统及相应特征,充分利用新标签信息。最后利用BCI竞赛2005的数据I证明LCTSSL算法的有效性。

A like-co-training semi-supervised learning algorithm for brain-computer interface is proposed aiming at improving the classification accuracy when a few of samples＇ labels have been known. Two kinds of features are extracted by supervised and unsupervised extractions respectively and two corresponding classifiers are trained. They cooperate with each other to enlarge the labeled set. Different standard for confidence is defined in different training stage. In order to improve the confidence of the enlarged labeled set and the robustness of the model of the pattern recognition, the unlabeled whose confidence is higher and predicted labels by the two classifiers are same will be labeled and make sure that the amounts of new labeled in each category are same. Both of classifiers, the supervised extractor and the corresponding features are updated each iteration for the purpose of absorbing new labels＇ information. At last, the applying on data set I of BCI competition 2005 demonstrated the validity of our pro-posed algorithm.