交互式机器教学与机器学习算法的研究

In recent years, machine learning has been intensively studied by the researchers all over the world, and has achieved great success in many situations. However, as an important component in a completed learning system, machine teaching has been largely ignored, which is unfavorable for the learning algorithm to obtain the satisfactory performance. Therefore, inspired by the cognitive process of humans, this project aims to fuse machine teaching and machine learning into a unified framework, so that the learning process is logically organized under the supervision of a teaching algorithm. Specifically, we treat the teaching algorithm and learning algorithm as “teacher” and “learner”, respectively, and in each learning round the teacher should choose the most suitable curriculum examples for the learner. After “learning” these examples, the learner should generate a learning feedback to assist the teacher to decide the optimized curriculum in the next round. By developing advanced machine teaching algorithms, establishing suitable learning feedback, and designing the effective unified framework, this project targets to improve the accuracy and robustness of machine learning algorithms, decrease their training time, and also apply the proposed interactive teaching and learning methodology to more practical problems.

近年来，各类机器学习算法已被国内外众多学者广泛地研究，并在各类应用中取得了丰硕的成果。然而，作为学习系统重要的另一面，机器教学的研究却一直被忽视，这就阻碍了学习算法取得更好的性能。因此，本项目受人类认知过程的启发，力图将机器教学和机器学习写进一个完整的学习系统框架，使得机器学习算法能够在机器教学算法的“指导”下进行有序、合理的学习，从而获得更好的效果。具体来说，我们将教学算法和学习算法分别视为“老师”和“学生”，在每轮教学过程中老师需要将最适合学生学习的课程样本“教授”给学生；学生在学完这些样本后，需要生成一个学习效果反馈以帮助老师在下一轮教学时制定最优的课程。本项目通过研究先进机器教学算法的开发、机器学习效果反馈的构建、整体融合框架的设计，希望能够提升机器学习算法的准确性、鲁棒性，降低学习算法的训练时间，并力图将所设计的交互式机器教学和机器学习算法应用于更多实际问题。

受人类认知过程的启发，本项目将机器教学和机器学习建模成一个完整的学习系统框架，使得机器学习算法能够在机器教学算法的“指导”下进行有序、合理的学习，从而获得更好的效果。具体而言，主要研究了先进机器教学算法的开发、机器学习效果反馈的构建、整体融合框架的设计，提升了机器学习算法的准确性、鲁棒性。所设计的交互式机器教学和机器学习算法被成功应用于图像分类、人脸识别、文本分类等诸多实际问题，并获得了目前最好的效果。研究成果发表在13篇机器学习领域国际顶级期刊或会议上，如TPAMI、TCYB、TNNLS、AAAI、IJCAI等，并申请了4项发明专利。受本项目资助，申请人获得了吴文俊人工智能优秀青年奖、中国科协托举计划等多项荣誉，证明了本课题的研究成果获得了广泛的关注和认可。