RI: SMALL: Inducing Answer Set Programs to Provide Accurate and Concise Explanation of Machine-learned Models

人工智能（AI）/机器学习作为一种重要的技术越来越突出，将对我们的经济，工业，社会和学术界产生重大影响。现代机器学习方法的一个主要问题是它们无法向人类用户解释它们的决定。统计机器学习方法产生的模型是优化问题（如风险最小化或数据似然最大化）的复杂代数解决方案。缺乏直观的描述使得用户很难理解、验证或信任管理模型的基本规则。此外，这些方法无法为它们为新数据样本计算的预测提供理由。因此，人们对所谓的可解释人工智能有着浓厚的研究兴趣。该项目将开发方法来捕获机器学习模型背后的逻辑，使模型可向用户解释。这将允许用户改进模型，并将增强用户对这些模型的信任。 归纳逻辑编程（ILP）是一种已建立的技术，用于找到人类可理解的机器学习模型的规则。学习的规则表示为逻辑程序或Horn子句。然而，由于缺乏否定作为失败，霍恩子句提供有限的表示和推理时，正在研究的领域的背景知识是不完整的表达。此外，ILP在假设规则没有例外的情况下学习规则。这导致数据中的异常和噪声彼此无法区分。通常，规则本身的例外遵循一种模式，这些例外可以（递归地）学习为默认理论。据推测，一个学习的程序，包括这样一个默认的理论描述的基础模型更准确。这个项目扩展了基于知识的，可扩展的ILP算法，学习默认理论作为给定背景知识以及正面和负面例子的答案集程序。这些答案集程序旨在捕捉学习模型的逻辑，为其决策提供理由，提高用户对模型的信任，发现模型中的任何偏见，并遵守外部要求，如政府法规。 该项目的目的是推进ILP研究的最新发展，并为机器学习和可解释AI的一般领域做出贡献。该项目的结果将是开源的，旨在使使用机器学习的行业能够更好地理解和信任他们使用的学习模型。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Artificial Intelligence (AI)/Machine Learning is gaining prominence as an important technology that will have significant impact on our economy, industry, society, and academia. A major problem with modern machine learning methods is their inability to explain their decisions to human users. Statistical machine learning methods produce models that are complex algebraic solutions to optimization problems such as risk minimization or data likelihood maximization. Lack of intuitive descriptions makes it hard for users to understand, verify or trust the underlying rules that govern the model. Also, these methods cannot produce a justification for a prediction they compute for a new data sample. As a result, there is significant research interest in what is termed as Explainable AI. This project will develop methods to capture the logic behind machine learning models, making the models explainable to users. This will allow users to improve the models and will enhance users' trust in these models. Inductive Logic Programming (ILP) is an established technique to find the rules underlying a machine-learned model that are comprehensible to humans. The rules learned are represented as logic programs or Horn clauses. However, due to lack of negation-as-failure, Horn clauses offer limited expressiveness for representation and reasoning when the background knowledge about the domain being studied is incomplete. Additionally, ILP learns rules under the assumption that there are no exceptions to them. This results in exceptions and noise in the data being indistinguishable from each other. Often, the exceptions to the rules themselves follow a pattern, and these exceptions can be (recursively) learned as a default theory. It is hypothesized that a learned program that includes such a default theory describes the underlying model more accurately. This project extends heuristics-based, scalable ILP algorithms that learn default theories as answer set programs given background knowledge as well as positive and negative examples. These answer-set programs aim to capture the logic underlying a learned model to provide justifications for its decisions and to improve users' trust in it, discover any biases in the model, and comply with outside requirements such as governmental regulations. The aim of this project is to advance the state-of-the-art in ILP research and to contribute to the general area of machine learning and explainable AI. Results of the project will be open-sourced, with the aim to enabling industries that make use of machine learning to develop better understanding of and trust in the learned models they use.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.