CAREER: Towards a New Synthesis of Statistical Learning and Logical Reasoning

在过去的十年里，人工智能(AI)领域发生了翻天覆地的变化，采取了更加以数据为中心的方法。机器学习的广泛成功引发了这样一个问题：哪些人工智能任务适合纯学习，哪些任务需要经典的符号推理，以及我们是否可以从这两种方法的更紧密集成中受益。该项目对这一问题进行了研究，并具体询问了传统人工智能中研究的自动推理和知识表示的思想与现代连接主义和统计机器学习的相关性。它汇集了几个不同领域的专业知识、技术、洞察力和优势，这些领域通常是单独研究的：逻辑推理、概率推理、知识表示、统计学习和深度学习。这项研究采取的统一视角有可能对更广泛的人工智能领域产生变革，并对我们如何看待学习和推理之间的相互作用产生持久的影响。这项研究将使人工智能能够解决更大类别的问题，从而使其更加有效。更有能力的人工智能和机器学习方法将产生重大的科学后果，并在社会各个领域产生广泛影响，包括医疗保健、制造业、商业、金融、娱乐等。该项目通过综合研究和教育活动帮助传达这种新的理解。更具体地说，当前的推理范式不能充分利用现有数据，而且往往很脆弱，而学习范式往往无法回答超出其明确培训的一项任务的问题。找到学习和推理的综合体允许学习的表示可以被推理，甚至在学习过程中使用推理和逻辑，以加强基本的不变量和世界的知识。该项目围绕三个研究方向展开。第一个推力是开发概率和逻辑电路作为一种新的机器学习模型，该模型同时易于学习、表达，并具有允许易于处理的推理和学习的优雅特性。第二个重点着眼于关于分类器和生成世界模型的更高级的推理任务，例如在特征缺失时进行预期预测，或者对解释分类器的充分条件进行推理。具体地说，对分类器进行推理的能力，在我们的人工智能系统部署时建立了更多的信任，并有助于更好地了解它们的限制。第三个重点是研究如何将关于连续变量和算术的逻辑推理用于概率推理和统计学习。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Over the past decade, the field of artificial intelligence (AI) has evolved drastically, taking a more data-centric approach. The widespread success of machine learning raises the question of which AI tasks are amenable to pure learning, which tasks require classical symbolic reasoning, and whether we can benefit from a tighter integration of both approaches. The project studies this question, and concretely asks how ideas about automated reasoning and knowledge representation studied in traditional artificial intelligence are relevant to modern connectionist and statistical machine learning. It brings together expertise, techniques, insights, and strengths from several disparate fields that are usually studied in isolation: logical reasoning, probabilistic reasoning, knowledge representation, statistical learning, and deep learning. The unified perspective taken by this research has the potential to be transformative for the broader AI field and have a lasting impact on how we perceive the interaction between learning and reasoning. The research will make AI more effective by allow it to address a larger class of problems. More capable AI and machine learning methods will have significant scientific consequences, and broad impact in all segments of society, including healthcare, manufacturing, commerce, finance, entertainment, among others. The project helps convey this new understanding through integrated research and educational activities.More specifically, current reasoning paradigms are not able to fully exploit available data and are often brittle, while learning paradigms are often incapable of answering questions beyond the one task they were explicitly trained for. Finding a synthesis of learning and reasoning allows for learned representations that can be reasoned about, and even using reasoning and logic during learning, to enforce basic invariants and knowledge of the world. This project is structured along three research thrusts. The first thrust is to develop probabilistic and logistic circuits as a new machine learning model that simultaneously easy to learn, expressive, and has elegant properties that allow for tractable reasoning and learning. The second thrust looks at more advanced reasoning tasks about classifiers and generative world models, such as taking expected predictions when features are missing, or reasoning about sufficient conditions to explain classifiers. The ability to reason about classifiers, specifically, builds more trust in our AI systems as they are deployed, and helps to better understand their limitations. The third thrust studies how logical reasoning about continuous variables and arithmetic is used for probabilistic reasoning and statistical learning.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.