CAREER: Interpretable and Robust Machine Learning Models: Analysis and Algorithms

虽然机器学习在不同领域的影响不断增加——从推荐系统到算法交易，从医学成像诊断到分子生物学——但这些日益复杂的模型的一些局限性代表了它们安全和负责任的部署的重要缺点。这些限制之一是这些预测因素缺乏可解释性，因此难以忠实地确定给定输入中最相关部分在产生某种输出中的作用。另一个限制是它们的脆弱性，因为即使输入的扰动很小，这些输出也可能非常不稳定。这些问题可能会危及现代机器学习工具在敏感领域（如医学成像）的安全部署。本项目将开发正式的方法和算法来减轻这些缺点。该CAREER项目将开发一个通用框架，以稳健和可验证的方式解释复杂的预测因素。特别是，本项目将首先定义局部特征重要性的新概念，并开发正确有效的方法来估计它们。这些定义将根据局部条件独立性测试给出，同时对预测函数进行最小的假设，以及对语义上重要概念的扩展。其次，本项目将提出并分析算法来证明预测模型在流形上的局部稳定性，并保证模型解释的稳定性和鲁棒性。本项目衍生的方法将在一系列医学成像问题上进行评估，包括胸部x光和计算机断层扫描。此外，该项目将开展一项全面的教育和推广计划，致力于增加STEM中少数民族学生的代表性，包括通过约翰霍普金斯大学教育推广中心进行K-12和社区推广，战略本科和研究生研究项目，以及向科学界和公众广泛传播，以及其他举措。该奖项反映了美国国家科学基金会的法定使命，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。

While the impact of machine learning continues to increase in different areas---from recommendation systems to algorithmic trading, and from medical imaging diagnosis to molecular biology---some limitations of these increasingly complex models represent important shortcomings for their safe and responsible deployment. One of these limitations is the lack of interpretability of these predictors, making it difficult to faithfully determine the role of the most relevant portions of a given input in producing a certain output. Another limitation is their brittleness, as these outputs can also be remarkably unstable even to very small perturbations of the inputs. These problems can compromise the safe deployment of modern machine learning tools in sensitive domains, such as medical imaging. This project will develop formal methods and algorithms to alleviate these shortcomings.This CAREER project will develop a general framework to interpret complex predictors in a robust and certifiable manner. In particular, this project will first define new notions of local feature importance and develop correct and efficient methods to estimate them. These definitions will be given in terms of local conditional independence tests while making minimal assumptions about the prediction functions, as well as extensions to semantically important concepts. Second, this project will propose and analyze algorithms to certify the stability of predictive models locally on manifolds, as well as guarantee the stability and robustness of model interpretations. The methods derived from this project will be evaluated on a series of medical imaging problems that include chest X-rays and computed tomography. In addition, this project will carry out a holistic educational and outreach program dedicated to increasing the representation of minority students in STEM, including K-12 and community outreach through the Johns Hopkins Center for Educational Outreach, strategic undergraduate and graduate research projects, and broad dissemination to both the scientific community and the general public, among other initiatives.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.