Interactive online learning

人类通过学习获得了大部分先进的感知和表演能力。机器学习领域的想法是赋予计算机类似的学习能力，以便通过学习，它们可以克服由于我们缺乏对如何解决复杂任务的理解而造成的限制。机器学习算法的应用比比皆是。在我们的日常生活中，机器学习算法通过允许我们对手机通话、过滤电子邮件垃圾邮件、翻译文档或识别人脸来帮助我们，无论是出于安全原因还是在社交网络上给我们的朋友贴上标签。 以上是一次性预测问题的例子，在这种情况下，决策的效果不会产生长期影响。然而，在许多应用程序中，情况并非如此。例如，决定将什么内容放在门户网站的首页将限制收到的有关网站访问者偏好的信息：简单地说，没有关于未显示的内容的反馈。因此，决策者必须对因其决策而丢失信息的影响进行推理。在其他情况下，比如选择职业，今天所做的决定会影响未来出现的机会。我的研究重点是探索如何在这些交互场景中使用机器学习。 我的目标是构建能够处理比以前可行的更大规模和更复杂的问题的算法。我采用的策略是首先研究特殊情况，因为这使我能够孤立地识别和解决基本问题，最终导致能够解决更复杂问题的更高效的算法。例如，最近，使用这项技术，我在开发算法方面取得了进展，这些算法以一种可证明是最优的方式明确地推理“信息的价值”。这项工作有望通过将当前机器学习方法的适用范围扩展到无处不在的交互学习环境来实现广泛的影响。

Humans acquire most of their advanced sensing and acting capabilities by learning. The idea of the field of machine learning is to give computers similar learning capabilities so that through learning they can overcome the limitations imposed by our lack of understanding of how to solve complex tasks. Applications of machine learning algorithms abound. In our daily lives, machine learning algorithms help us by allowing us to speak to our phones, by filtering out e-mail spam, by translating documents, or by recognizing faces, either for security reasons or to label our friends on social networks. The above are examples of one-shot prediction problems, where the effect of a decision does not have long term impact. However, in many applications this is not the case. For example, the decision of what content to put on the frontpage of a web portal will limit the information received about the preferences of the website's visitors: Simply, no feedback will be available about content that is not shown. Thus, the decision maker has to reason about the impact of losing information as a result of his decisions. In other situations, such as choosing a profession, decisions made today influence what opportunities arise in the future. My research focuses on exploring how machine learning can be used in these interactive scenarios. My goal is to build algorithms that can handle problems of greater size and complexity than previously feasible. The strategy I employ is to study special cases first, as this allows me to identify and address fundamental issues in isolation, which eventually leads to more efficient algorithms capable of addressing more complex problems. For example, recently, using this technique, I made progress on developing algorithms that explicitly reason about the "value of information" in a provably optimal manner. This work is expected to achieve broad impact by extending the scope of applicability of current machine learning methods to the ubiquitous interactive learning settings.