RI:Small:Investigating techniques that couple Markov Logic and Deep Learning with applications to discovering strategies to improve STEM learning

该项目的目标是开发新的技术，以集成人工智能(AI)中不同但互补的方法。这项研究结合了深度神经网络(DNN)和马尔可夫逻辑网络(MLN)的优点，以解决这些技术在单独使用时的关键缺点。特别是，拟议的工作将解决DNN在学习模型时利用背景知识方面的限制。DNN通常不显式地利用背景知识的事实经常导致模型过度拟合训练数据并且在新数据集上泛化得很差。另一方面，马尔可夫逻辑网络(MLN)等统计关系模型显式地编码复杂的背景知识，但缺乏与基于DNN的方法一样的可扩展性和准确性的推理和学习能力。该项目将开发新的技术，其中MLN向DNN提供特定于任务的背景知识，帮助DNN学习更多可推广的模型。此外，该项目将应用这些新技术来显著改善STEM主题自适应教学系统(AISS)中的个性化学习。该项目将产生i)可被广泛的应用领域使用的学习和推理的通用开放源码软件，以及ii)基于AIS的学习的核心任务的特定模型(例如，推断学生的问题解决策略)，该模型可以显著提高AISS的适应能力，从而使学生更好地参与和学习。该项目将影响多个社区，包括机器学习、人工智能、教育中的人工智能和教育数据挖掘。我们的工作成果将通过在顶级会议和期刊上的出版物、演示文稿、网站、社交媒体以及研究人员和从业人员的培训材料广泛传播。将背景知识结合到DNN中的现有方法使用贝叶斯框架来做到这一点，其中先验类型通常很简单，以确保贝叶斯推理的易处理性。该项目的主要技术贡献是通过开发包含以MLN形式指定的丰富关系知识的DNN模型来解决这一限制。为此，该项目将：i)开发将MLN分布中的对称性(或可交换性)编码为子符号嵌入的新表示，ii)通过利用MLN隐含指定的变量的可交换性来开发用于关系数据的高效的基于DNN的学习算法，以及iii)使用生成性对抗性网络来开发可解释的生成模型，利用MLN指定的对称性来遍历分布中的不同模式。将作为该项目的一部分开发的AIS任务将使用大规模数据集，从而令人信服地展示所提议的模型在现实世界问题中的可扩展性。此外，为AIS任务开发的模型将帮助我们更好地了解学生的需求和学习过程，这反过来可以为STEM主题的先进教育技术的改进提供信息，并帮助验证和完善人类学习理论。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。

The goal of this project is to develop novel techniques to integrate different but complementary approaches in artificial intelligence (AI). This research combines the strengths of Deep Neural Networks (DNNs) and Markov Logic Networks (MLNs) to address key shortcomings of those techniques when used by themselves. In particular, the proposed work will address the limitation of DNNs with respect to utilizing background knowledge in learning a model. The fact that DNNs typically do not utilize background knowledge explicitly often results in models that over-fit the training data and generalize poorly on new datasets. On the other hand, statistical relational models such as Markov Logic Networks (MLNs) encode complex background knowledge explicitly but lack inference and learning capabilities that are as scalable and accurate as DNN-based methods. The project will develop novel techniques in which MLNs provide the DNN with task-specific background knowledge which helps the DNN to learn more generalizable models. Further, this project will apply these novel techniques to significantly improve personalized learning in adaptive instructional systems (AISs) for STEM topics. The project will yield i) general-purpose open-source software for learning and inference that can be used by a broad range of application domains and ii) specific models for core tasks in AIS-based learning (e.g. inferring student problem-solving strategies) that can significantly improve the adaptive capabilities of AISs which results in better student engagement and learning. The project will impact a number of communities including machine learning, artificial intelligence, artificial intelligence in education, and educational data mining. The outcomes of our work will be widely disseminated through publications in top conferences and journals, presentations, a website, social media, and training materials for researchers and practitioners. Existing approaches that incorporate background knowledge into DNNs do so using a Bayesian framework where the types of priors are typically simple to ensure tractability of Bayesian inference. The main technical contribution of this project is to address this limitation by developing DNN models that incorporate rich relational knowledge specified in the form of an MLN. To do this, the project will i) develop new representations that encode symmetries (or exchangeability) in the MLN distribution as sub-symbolic embeddings, ii) develop efficient DNN-based learning algorithms for relational data by exploiting exchangeability of variables specified implicitly by the MLN and iii) develop interpretable generative models using Generative Adversarial Networks utilizing symmetries specified by the MLN to traverse across diverse modes in the distribution. The AIS tasks that will be developed as part of this project will use large-scale datasets and thus convincingly demonstrate the scalability of the proposed models in real-world problems. Further, the models developed for the AIS tasks will help us better understand student needs and learning processes which in turn can inform improvements of advanced educational technologies for STEM topics and help validate and refine human learning theories.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.