A unifying framework for integrating domain knowledge into machine learning algorithms for multidisciplinary industrial applications

尽管监督学习方法在预测分析中表现优异，但迄今为止，它们只能从标记数据中学习。在这个过程中，人类的作用被限制在“仅仅是标签者”，专家的知识没有得到充分利用。这项研究设想了一个通用框架的发展，将专家的意见代表和整合到学习系统中，如果可能的话，将他们宝贵的知识转移到现实生活中的应用决策过程中。在接下来的五年里，我们计划开发一个新的统一框架，以有效的方式利用人类的建议来促进机器学习（ML）算法，我们打算为工业应用开发人类指导学习的新解决方案。拟议研究的长期目标包括深入调查和开发新颖，自动化，高效的实时ML算法，能够通过利用有监督和无监督方法的领域知识为工业问题提供解决方案。创新：将专家的知识整合到嘈杂、结构化领域的机器学习算法中的新方法将被开发出来，以加速学习有效的模型，到目前为止，人类仅被用作标注者。我们计划开发一个自然的框架，允许专家对他们的知识进行编码，既可以作为关于领域的一般建议，也可以作为关于特定示例的具体建议。拟议的框架将通过优先、基于成本函数、质量限制和特权信息获取不同类型的咨询意见，从而提供通用性。实际应用：在实践中，本研究有可能产生重大影响：(i)医学成像和法医等领域的工业应用，这些领域可以受益于图像分割和配准技术，并且可以通过有效地使用深度学习来利用人类的建议；（二）焊接行业，通过加快焊接顺序优化和减少结构变形；（iii）制造业，通过自动化操作，如钉孔插入任务；（四）机器人行业，通过为更高自由度的机器人逆运动学问题提供解决方案。高素质人才培养：拟议的研究计划旨在培养学生和培养在机器学习和人工智能（AI）方面具有独特专业知识的数据科学家，使他们成为机器学习、人工智能和计算机视觉领域的世界级专家。

Though supervised learning methods demonstrate superior performance in predictive analytics, to date, they learn only from labeled data. The role of human beings in the process has been restricted to "mere labeler," and the knowledge of experts is not being fully utilized. This research envisages the development of a general framework to represent and integrate experts' opinions into learning systems, and, if possible, to transfer their invaluable knowledge for decision-making processes in real-life applications. Over the next five years, we plan to work on developing a novel unifying framework for utilizing human advice in an efficient manner to facilitate machine learning (ML) algorithms and we intend to develop novel solutions to human-guided learning for industrial applications. The long-term goals of the proposed research include deep investigation and development of novel, automated, efficient real-time ML algorithms capable of providing solutions to industrial problems by exploiting domain knowledge in both supervised and unsupervised methods. Innovation: Novel approaches for incorporating experts' knowledge into ML algorithms in noisy, structured domains will be developed to accelerate learning effective models in which, so far, humans have been merely used as labelers. We plan to develop a natural framework which will allow experts to encode their knowledge, both as general advice about the domain and as specific advice about particular examples. The proposed framework will offer generality by capturing different types of advice through preferential, cost function-based, qualitative constraints, as well as privileged information. Practical Applications: In practice, this research has the potential to significantly impact: (i)industrial applications in fields such as medical imaging and forensics, which can benefit from image segmentation and registration techniques and in which human advice can be exploited by using deep learning in an effective manner; (ii)the welding industry, by accelerating welding sequence optimization and reducing structural deformations; (iii)the manufacturing industry, by automating operations such as peg-hole insertion tasks; and (iv)the robotics industry, by providing solutions to robot inverse kinematics problems for higher degrees of freedom. Highly Qualified Personnel Training: The proposed research program intends to train students and produce data scientists with unique expertise in ML and artificial intelligence (AI), which will enable them to be world-class experts in the fields of ML, AI, and computer vision.