Development of Multitask Pattern Recognition Model with Knowledge Transfer of Feature Space and Application to Person Identification

在顺序地学习具有某种相关性的多个模式识别任务的环境中，已知的是，通过使用从一个任务到另一个任务的“知识转移”，即使利用少量的训练数据也能有效地进行学习。在研究项目中，我们开发了一个多任务学习算法与有效的知识转移机制，一个有用的特征空间是学习增量在一个有效的方式通过转移一部分以前学到的知识到一个未知的任务。提出的多任务学习算法实现为一个人识别系统，使用人脸图像和该系统的有效性进行了验证。

In the environments where multiple pattern recognition tasks with some relatedness are learned sequentially, it is known that the learning is conducted efficiently even with a small number of training data by using "knowledge transfer" from one task to another. In the research project, we developed a multitask learning algorithm with an efficient knowledge transfer mechanism where a useful feature space is learned incrementally in an efficient way by transferring a part of previous learned knowledge to an unknown task. The proposed multitask learning algorithm is implemented as a person identification system using face images and the effectiveness of this system is verified.