CAREER: From Rare Events to Competitive Learning Algorithms

职业：从罕见事件到竞争性学习算法

• 批准号: 2146334
• 负责人: Mesrob Ohannessian
• 金额: $ 54.57万
• 依托单位: University of Illinois at Chicago
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-07-01 至 2027-06-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2146334&HistoricalAwards=false
• 关键词: CAREER; Rare; Events; Competitive; Learning

A sense of optimism underlies the mass adoption of data-driven algorithms. Such algorithms will be used not only to match current scientific and engineered solutions but also, given enough data, to rival years of future effort. This project proposes framing optimism in terms of competitive learning, where a single algorithm performs as well as a family of bespoke algorithms, each for a specific true nature. Thus, even without knowledge of the truth, a competitive algorithm is guaranteed to deliver the best feasible performance in each case. This approach shows that much more is possible than the pessimistic outlook of judging performance against a worst-case nature. It can also help better understand the limits of (human) discovery, based on how amenable nature is to being discovered. By cutting right to the heart of the philosophy of acquiring knowledge from observation, this work weaves research with data science education efforts at several levels, to help nurture and train the next generation of data scientists and engineers, with active outreach and engagement of traditionally underserved students.The project fully explores the competitive perspective for contextual distribution learning, a problem that permeates data science and machine learning, with natural language modeling as a specific use case. Here, the goal is to demonstrate an algorithm that learns to predict, by building a stochastic matrix or tensor from data. It is successful if its performance rivals that of a rich array of specialized algorithms aware of potential underlying structures, such as rank, sparsity, or low manifold dimension, despite no single algorithm being worst-case optimal. The larger premise of the project is that the principles that enable competitiveness are fundamental and cut across multiple domains. These include: (i) the back-off principle, which demarcates between abundant and rare data regions and applies structure only in the latter, where it is needed, achieving competitiveness through nimbleness, (ii) the empirical-Bayes principle, which offers a mechanism to share data across rarely represented regions, allowing them to help each other, and (iii) tail structures, which ground the first two principles in a tractable framework and can be used to establish fundamental limits, by characterizing conditions that are necessary and sufficient for competitiveness. By rigorously establishing the foundations of these principles, the goal of the project is to streamline the design of competitive algorithms that simultaneously find the truth of nature and adapt to it.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.

乐观主义是数据驱动算法大规模采用的基础。这些算法不仅将用于匹配当前的科学和工程解决方案，而且如果有足够的数据，还将与未来多年的努力相媲美。该项目提出了在竞争性学习方面的框架乐观主义，其中单个算法的性能以及一系列定制算法，每个算法都具有特定的真实性质。因此，即使不知道真相，竞争性算法也能保证在每种情况下提供最佳可行性能。这种方法表明，比根据最坏情况的性质来判断业绩的悲观看法，还有更多的可能性。它还可以帮助更好地理解（人类）发现的局限性，基于自然如何被发现。通过切入从观察中获取知识的哲学核心，这项工作将研究与多个层面的数据科学教育工作结合起来，以帮助培养和培训下一代数据科学家和工程师，并积极推广和参与传统上服务不足的学生。该项目充分探索了上下文分布学习的竞争前景，这是一个渗透到数据科学和机器学习中的问题，自然语言建模是一个特定的用例。在这里，我们的目标是展示一种算法，通过从数据中构建随机矩阵或张量来学习预测。它是成功的，如果它的性能竞争对手的一系列丰富的专门算法知道潜在的底层结构，如排名，稀疏性，或低流形维数，尽管没有一个算法是最坏情况下的最佳。该项目的大前提是，使竞争力的原则是根本的，跨越多个领域。其中包括：（i）回退原则，其在丰富和稀有数据区域之间划界，并且仅在需要时在后者中应用结构，通过灵活性实现竞争力，（ii）概率贝叶斯原则，其提供了一种机制，以在很少代表的区域之间共享数据，允许它们互相帮助，以及（iii）尾部结构，这两项原则将前两项原则建立在一个易于处理的框架中，并可用于确定基本限制，确定竞争力所必需和充分的条件。通过严格建立这些原则的基础，该项目的目标是简化竞争性算法的设计，同时发现自然的真相并适应它。该奖项反映了NSF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而被认为值得支持。