CAREER: Machine Learning through the Lens of Economics (And Vice Versa)

机器学习(ML)是利用数据和计算资源来获得在存在不确定性的情况下运行良好的预测和决策算法的研究。用于设计和研究ML算法的技术通常涉及来自概率、统计和优化的概念和工具；另一方面，经济学的语言明显缺失。在ML研究文献中，很少遇到边际价格、效用、均衡、风险规避等术语。这一差距是巨大的，掩盖了这样一个现实：人们对机器学习的广泛兴趣，以及它作为一个研究领域的突然增长，可以归因于它在社会许多阶层产生经济价值的潜力。这个NSF职业项目推进了机器学习与微观经济理论和金融领域之间已经出现的关系。这将从开发数学工具开始，这些工具使学习理论对象和经济抽象之间能够实现语义对应。例如，该项目表明，许多算法可以被视为实施市场经济，其中学习参数与价格相关联，参数更新被视为交易，并且在某些条件下，学习的假设可以被提取为市场清算价格均衡。除了发展这种联系，该项目研究还提出了一些有趣的问题，并探索了几个令人惊讶和新颖的应用程序，这些应用程序对更广泛的计算机科学有好处。在新的理论联系产生的几个这样的应用中：1.为学习和估计任务开发新的分布式计算模型：经济学的镜头为分散数据重点任务的稳健和有效的模型提供了新的见解。通过涉及金融支付方案的协作机制，设计众包和劳动力分散的新技术：这建立在亚马逊的机械土耳其以及Netflix奖和预测挑战公司Kaggle等平台的成功基础上。为数据经纪和经济高效的学习开发一个以市场为导向的模型：随着信息在市场环境中的交易日益频繁，我们的目标是回答诸如“一个数据单位的边际价值是多少？”等问题。该项目还将开发密歇根预测团队，这是一个专注于数据科学的项目，旨在制定和解决密歇根大学内外发展起来的预测和学习挑战。该团队主要针对本科生和研究生导师，团队有很强的跨学科重点。

Machine Learning (ML) is the study of leveraging data and computational resources to obtain prediction and decision-making algorithms that function well in the presence of uncertainty. The techniques employed to design and study ML algorithms typically involve concepts and tools from probability, statistics, and optimization; the language of economics, on the other hand, is conspicuously absent. It is rare to encounter terms such as marginal price, utility, equilibrium, risk aversion, and such, in the ML research literature. This gap is significant and belies the reality that the broad interest in Machine Learning, and its sudden growth spurt as a research field, can be ascribed to its potential for generating economic value across many segments of society. This NSF CAREER projectadvances an already-emerging relationship between Machine Learning and the fields of microeconomic theory and finance. This will begin with the development of mathematical tools that enable a semantic correspondence between learning-theoretic objects and economic abstractions. For example, the project shows that many algorithms can be viewed as implementing a market economy, where learning parameters are associated with prices, parameter updates are viewed as transactions, and under certain conditions learned hypotheses can be extracted as market-clearing price equilibria. In addition to developing this link, the project research raises a number of intriguing questions and explores several surprising and novel applications with benefits to computer science more broadly. Among several such applications stemming from the new theoretical connections are:1. Developing new models for distributed computing for learning and estimation tasks: The economic lens gives new insights into a robust and effective model for decentralization of data-focused tasks.2. Designing new techniques for crowdsourcing and labor decentralization via collaborative mechanisms involving financial payment schemes: This builds off of the success of platforms like Amazon's Mechanical Turk as well as the Netflix Prize and the prediction challenge company Kaggle.3. Developing a market-oriented model for data brokerage and financially-efficient learning: As information is increasingly traded in market environments, we aim to answer questions such as "what is the marginal value of a unit of data?"The project will also develop the Michigan Prediction Team, a data-science focused program for formulating and solving prediction and learning challenges that develop from across the University of Michigan as well as externally. The group primarily targets undergraduates with graduate student mentors, and Team has a strong interdisciplinary focus.