CAREER: Generative Models for Targeted Domain Interpretability with Applications to Healthcare

人工智能和机器学习即将在自动驾驶、个性化新闻馈送和治疗推荐系统等特征不佳的环境中部署，这创造了对解释其决策的机器学习系统的迫切需求。可解释性帮助人类专家确定经过技术目标函数训练的机器学习系统是否具有合理的输出，尽管未建模的未知数。例如，临床决策支持系统永远不会知道患者的所有病史，也可能不知道特定患者愿意容忍许多副作用中的哪一种。因此，一个重要的挑战是如何设计出既能准确预测又能提供解释的机器学习系统。在这一广泛的挑战中，这项工作开发了面向领域的可解释性技术，找到与决策相关的高维数据摘要。拟议的工作重点是医疗保健应用，其中可解释的模型对安全至关重要。然而，该项目的目标是产生适用于一系列科学和社会领域的基础学习算法。开发的方法将在脓毒症、抑郁症和自闭症谱系障碍的个性化治疗建议和预后方面的实际问题上进行测试。因此，这项工作的成功完成将对可解释机器学习和临床科学产生影响。在该项目过程中开发的所有软件将免费共享。拟议工作的教育部分将教育早期小学生了解统计学在医学中的影响，并教育政策制定者和法律学者如何在诸如临床决策支持系统等机器学习的背景下规范解释权。PiDoshi-Velez还邀请了高中生、本科生、女性和她实验室中服务不足地区的研究人员。拟议的工作解决了科学环境中常见的一个特定挑战：针对领域的可解释性。在许多科学领域，领域专家使用非监督生成模型来理解数据中的模式，但随着数据维度的增加，数据中最显著的模式可能与特定调查无关。例如，精神病学家可能会在他的患者队列数据中发现最强的信号来自糖尿病和心脏病，这可能与选择抑郁症的治疗方法无关。拟议的工作利用协同效应来解释数据中与领域相关的模式，并在与领域相关的任务中很好地执行，以实现针对领域的可解释性。它定义了一种针对领域的可解释性的任务受限方法，并开发了基本的推理技术，开发了对顺序决策的扩展，并定义了扩展以在保持可解释性的同时提高下游任务的性能。虽然在使无监督学习模型对下游任务也有用方面有大量的工作，但这些方法都没有真正管理提供数据解释和任务性能之间的权衡。本文针对这些不足，提出了针对领域的可解释性和任务绩效协同目标，并提出了一些创新之处。创新包括将现有丰富的推理文献、传统的无监督模型与现代推理技术相结合，并直接搜索与下游任务相关的维度或模式。

The imminent deployment of AI and machine learning in poorly-characterized settings such as autonomous driving, personalized news feeds, and treatment recommendation systems has created an urgent need for machine learning systems that explain their decisions. Interpretability helps human experts ascertain whether machine learning systems, trained on technical objective functions, have sensible outputs despite unmodeled unknowns. For example, a clinical decision support system will never know all of a patient's history, nor may it know which of many side effects a specific patient is willing to tolerate. An important challenge, then, is how to design machine learning systems that both predict well and provide explanation. Within this broad challenge, this work develops techniques for domain-targeted interpretability, finding summaries of high-dimensional data that are relevant for making decisions. The proposed work focuses on healthcare applications, where interpretable models are essential to safety. However, the project aims to produce foundational learning algorithms applicable to a range of scientific and social domains. The developed methods will be tested on real problems in personalizing treatment recommendations and prognoses for sepsis, depression, and autism spectrum disorder. Thus, the successful completion of the work will impact both interpretable machine learning and clinical science. All software developed in the course of the project will be freely shared. The educational component of the proposed work will educate early elementary students about the impact of statistics in medicine and educate policy-makers and legal scholars on how a right to explanation might be regulated in the context of machine learning, such as clinical decision support systems. PI Doshi-Velez also engages high school students, undergraduates, women, and researchers from underserved areas in her lab.The proposed work addresses a specific challenge common in scientific settings: domain-targeted interpretability. In many scientific domains, unsupervised generative models are used by domain experts to understand patterns in the data, but as the dimensionality of data grow, the most salient patterns in the data may not be relevant for the specific investigation. For example, a psychiatrist may find the strongest signals in the data from his patient cohort come from diabetes and heart disease, which may not be relevant for choosing therapies for depression. The proposed work leverages synergies in explaining domain-relevant patterns in the data and performing well on domain-relevant tasks to achieve domain-targeted interpretability. It defines a task-constrained approach to domain-targeted interpretability and develops essential inference techniques, develops extensions to sequential decision making, and defines extensions to improve downstream task performance while retaining interpretabilty. While there is a large body of work on making unsupervised learning models also useful for downstream tasks, none of these approaches truly manage the trade-offs between providing an interpretation of data and task performance. The proposed work addresses these shortcomings to make domain-targeted interpretability and task performance synergistic goals, and proposes a number of innovations to solve the proposed objective. Innovations include combining an existing rich literature on inference traditional unsupervised models with modern inference techniques and directly searching for dimensions or patterns relevant to the downstream task.