ITR - (EVS+NHS) - (dmc + int): Knowledge Infusion

这项拟议的研究的目标是扩大机器学习技术的覆盖范围，使其能够使计算机执行比目前更广泛的任务。具体地说，目标是使机器能够从数据中提取知识，并将其转化为一种形式，以便对其进行稳健的推理。目前，可以在其上进行大规模推理的表示通常是编程的，结果存在脆性--它们在不可预见的情况下表现不佳。在所提出的方法中，称为知识灌输，目标是获得将通过编程和学习相结合进行推理的规则，并有一个连续的学习和对照环境的过程，以确保规则是可靠的。目标是处理未公理的或常识性知识，这些知识包括人类每天处理的大量知识，如体现在言语或文本中的知识，因为这些知识往往充满了不一致、模棱两可和错误。这可以与已知的可公理化的知识区分开来，例如大多数数学性质的知识。一般说来，公理化的知识可以很容易地编程，计算机通常可以充分利用这种知识，直到手头问题的任何固有计算复杂性限制。研究的最核心目标是开发实现知识灌输的算法，即使对于非常大的数据集，算法也是计算高效和有效的。同样重要的是确定这一现象的根本限制是什么。使用的技术将来自理论计算机科学，并将根据需要在大型数据集上进行实验。其目标是能够大规模地向机器注入关于世界的常识知识，使机器能够以受控的健壮性水平与之进行推理。这一努力的成功可以预期在几乎所有计算领域都有应用，这些领域要么涉及人与计算机的交互，要么涉及对由人类产生的或与人类有关的数据的计算。因此，与国家优先领域电动汽车和国民保健服务以及技术重点领域INT和DMC有许多联系。更广泛的影响：如果研究成功，研究结果将有助于提高计算机处理有关世界的常识或非公理信息的效率。这将把计算机的用途扩展到新的领域，并有助于繁荣(EVS)。它还将使大型数据集能够自动分析，具有比迄今(NHS)更大的功能。

The goal of the proposed research is to extend the reach of machine learning technology so that it can enable computers to perform a broader span of tasks than currently. In particular the goal is to enable machines to extract knowledge from data into a form such that robust reasoning can be done on it. Currently representations on which reasoning can be done on a large scale are typically programmed and the results suffer from brittleness - they do not behave well in unforeseen situations. In the proposed approach, which is called knowledge infusion, the goal is to acquire the rules on which reasoning will be done by a combination of programming and learning, and to have a continuous process of learning and checking against an environment to ensure that the rules are reliable. The goal is to handle unaxiomatized or commonsense knowledge, which encompasses the bulk of knowledge that humans handle everyday, as embodied in speech or text, replete as these often are with inconsistencies, ambiguities and errors. This can be distinguished from knowledge that is known to be axiomatizable, such as most knowledge of a mathematical nature. Axiomatized knowledge, in general, can be easily programmed, and computers can usually fully exploit such knowledge up to any inherent computational complexity limitations of the problem at hand. The most central aims of the research are the development of algorithms that realize knowledge infusion and are computationally efficient and effective even for very large datasets. Also central is the identification of what the fundamental limits of the phenomenon are. The techniques used will be from theoretical computer science, and experimentation on large datasets will be carried out as needed. The goal is to be able to infuse into machines commonsense knowledge about the world on a large scale and in a way such that the machines will be able to reason with it with a controlled level of robustness. Success in this endeavor can be expected to have applications in almost all areas of computing that involve either human interaction with a computer, or computation on data that was generated by or has reference to humans. Hence there are numerous connections with the national priority areas EVS and NHS, and with the technical focus areas int and dmc. Broader Impact: If successful the results of the research will help enhance the effectiveness of computers to handle commonsense or unaxiomatized information about the world. This would extend the usefulness of computers to new areas and contribute to prosperity (EVS). It would also enable large datasets to be analyzed automatically with greater functionality than hitherto (NHS).