Theoretical Foundations for Safe, Adaptive and Interpretable Machine Learning

提议的研究是在发展机器学习的理论基础领域。机器学习技术正被应用于各种日常生活系统中，对工业、科学和社会的影响越来越大。这包括推荐系统或个人设备等商业用途，也包括自动驾驶汽车、脑机接口或医疗援助等关键应用。由于机器学习实践目前取得了巨大的成功，并且学习方法被应用于大量的应用领域，因此对这些方法进行可靠和形式化的分析和理解变得越来越重要。目前许多最成功的学习方法（特别是基于深度学习的方法）都有一个“黑盒子”的本质。虽然他们在基于大量数据构建准确预测方面取得了巨大成功，但社会对数据驱动工具以不断扩大但不受控制的方式影响人们生活的不安感日益增强。事实上，我们对这些方法的工作原理和方式，影响结果的因素，可以为结果提供哪种类型的正确性保证，以及我们如何解释和传达机器学习系统做出的决定，知之甚少。

The proposed research is in the area of developing theoretical foundations to machine learning. Machine learning technologies are being used in various every-day life systems with an ever growing influence on industry, science and society. This includes commercial usage such as in recommendation systems or personal devices, but also critical applications such as autonomous vehicles, brain-machine interfaces or medical assistance. As machine learning practice currently enjoys enormous success and learning methods are employed in a vast number of application domains, it becomes increasingly crucial to develop solid and formal analysis and understanding of these methods. Many of the currently most successful learning methods (in particular deep learning based ones) have a “blackbox” nature. While they are becoming enormously successful at building accurate predictors based on large amounts of data, there is a growing societal sense of discomfort about data-driven tools affecting people's lives in ever expanding, yet uncontrolled, ways. We have, in fact, distressingly little understanding on why and how these methods work, what influences their outcomes, what type of correctness guarantees one can provide for their outcomes, and how we may interpret and communicate decisions made by a machine learned system to non-experts. The long-term goal of the proposed research program is to develop a better formal understanding of requirements and methods for safe and interpretable machine learning. To start addressing this challenge, over the next few years, we plan to explore the potential of interactive machine learning for this purpose. So far, research in interactive learning has mostly focused on how to design active label queries to efficiently achieve high statistical accuracy in supervised learning tasks. Here we aim to explore the potential of active feedback loops in learning process for three other objectives: (1) Safety as machine learning technologies enter critical applications, such as self-driving cars or medical devices, we need stronger guarantees of correct behavior than the statistical guarantees we currently rely on. (2) Adaptability machine learning systems are often trained in environments that are different from the one they are employed in. Thus, we need ways to let the system adapt to such changes. (3) Interpretability as machine learned systems are used for decision support directly affecting human individuals such as in the financial sector or in the judicial system, we need to develop ways to make these decisions comprehensible by humans. To this aim, we plan to develop novel protocols of user interaction with a learning system. Developing such from a foundational perspective will help building bridges between this fast developing technology and non-experts who are using or who are affected by machine learning.