CAREER: Learning Algorithms with Robustness and Efficiency Guarantees

职业：学习具有鲁棒性和效率保证的算法

• 批准号: 2144298
• 负责人: Ilias Diakonikolas
• 金额: $ 63.98万
• 依托单位: University of Wisconsin-Madison
• 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: 2022-03-01 至 2027-02-28
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2144298&HistoricalAwards=false
• 关键词: CAREER; Learning; Algorithms; Robustness; Efficiency

From achieving super-human performance in image classification to beating human champions in complex games, machine learning has seen enormous success in recent years. To unravel the full potential of machine learning and extend its reach, computer scientists seek to develop a comprehensive theory that explains the power and limitations of existing methods and offers avenues for improvement. This project aims to contribute in this direction by tackling two pressing challenges that limit the broader applicability of machine learning: (i) solving more complex tasks with multiple actions and decisions over time and (ii) learning from corrupted data. The goal is to study the performance and potential shortcomings of existing approaches through a unified framework and develop novel algorithms that are provably robust and efficient. In tandem with its research goals, the project incorporates the development of undergraduate and graduate courses at UW Madison, the training of graduate students, and research opportunities for undergraduates.In more detail, the project focuses on the themes of learning under noisy data and learning combinatorial algorithms from data.In the first theme, the investigator will build upon recent advances in dealing with noisy labels extending the results to more complex settings like multi-class classification, and will design more efficient methods that perform well in many practical situations. In the second theme, the investigator will study the use of Machine Learning to automatically develop efficient algorithms tailored to a specific application by bridging the areas of data-driven algorithm design and reinforcement learning. The focus in both themes will be on the statistical and computational complexity of the proposed methods. The project has the potential to offer a new paradigm of algorithm design through learning and to robustify machine learning systems enabling new application domains that involve large amounts of noise.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.

从在图像分类中实现超人表现到在复杂游戏中击败人类冠军，机器学习近年来取得了巨大成功。 为了充分发挥机器学习的潜力并扩大其影响范围，计算机科学家寻求开发一种全面的理论来解释现有方法的力量和局限性，并提供改进的途径。 该项目旨在通过解决限制机器学习更广泛应用的两个紧迫挑战来朝这个方向做出贡献：（i）随着时间的推移通过多种行动和决策解决更复杂的任务；（ii）从损坏的数据中学习。目标是通过统一的框架研究现有方法的性能和潜在缺点，并开发可证明稳健且高效的新颖算法。与其研究目标相一致，该项目结合了威斯康星大学麦迪逊分校本科生和研究生课程的开发、研究生的培训以及本科生的研究机会。更详细地说，该项目的主题是噪声数据下的学习和学习来自数据的组合算法。在第一个主题中，研究人员将基于处理噪声标签的最新进展，将结果扩展到更复杂的设置（例如多类分类），并将设计在许多实际情况下表现良好的更有效的方法。在第二个主题中，研究人员将研究如何使用机器学习，通过连接数据驱动算法设计和强化学习领域，自动开发针对特定应用量身定制的高效算法。这两个主题的重点都将集中在所提出方法的统计和计算复杂性上。该项目有潜力通过学习提供算法设计的新范式，并增强机器学习系统，从而实现涉及大量噪声的新应用领域。该奖项反映了 NSF 的法定使命，并通过使用基金会的评估进行评估，被认为值得支持。智力价值和更广泛的影响审查标准。

项目成果

Buying Information for Stochastic Optimization

购买随机优化信息

• 发表时间: 2023-06
• 期刊: 2023
• 作者: Ma, Mingchen;Tzamos, Christos
• 通讯作者: Tzamos, Christos

Optimizing Solution-Samplers for Combinatorial Problems: The Landscape of Policy-Gradient Methods

优化组合问题的解决方案采样器：策略梯度方法的概况

• 发表时间: 2023-11
• 期刊: 2023
• 作者: Caramanis, C;Fotakis, D;Kalavasis, A;Kontonis, V;Tzamos, C
• 通讯作者: Tzamos, C

Learning General Halfspaces with Adversarial Label Noise via Online Gradient Descent

通过在线梯度下降学习具有对抗性标签噪声的一般半空间

• 发表时间: 2022-01
• 期刊: 39th International Conference on Machine Learning (ICML
• 作者: Diakonikolas, Ilias;Kontonis, Vasilis;Tzamos, Christos;Zarifis, Nikos
• 通讯作者: Zarifis, Nikos

Approximating Pandora's Box with Correlations

用相关性逼近潘多拉魔盒

• 发表时间: 2023-06
• 期刊: 2023
• 作者: Chawla, Shuchi;Gergatsouli, Evangelia;McMahan, Jeremy;Tzamos, Christos
• 通讯作者: Tzamos, Christos

Weitzman's Rule for Pandora's Box with Correlations

韦茨曼潘多拉魔盒相关规则

• 发表时间: 2023-12
• 期刊: 2023
• 作者: Gergatsouli, Evangelia;Tzamos, Christos
• 通讯作者: Tzamos, Christos