RI: Small: Training Modularized Learning Systems

机器学习系统在现代生活中变得无处不在，也越来越复杂。许多设备，如手机，持续运行几十个预测模型，这些模型不仅从用户那里接收输入，也从彼此那里接收输入。考虑多重交互学习系统的意外挑战的一种方法是考虑人类如何在人际关系中互动，甚至是政府如何在国际争端中相互接触。这种情况涉及难以预测的动态，其中引入少量信息或对策略进行微小更改可能会产生截然不同的结果。该项目旨在从算法的角度理解这些相互作用的动态，着眼于设计模块化学习系统，在该系统中，实施者可以确定训练的动态将达到所需的解决方案。这项工作将大大增加在现实世界中应用学习系统的任务和挑战的范围，并将对人工智能如何与社会互动产生强烈影响。该项目首先关注博弈论，并建立在解决所谓最小-最大问题的一些经典和最新结果的基础上，其中人们希望找到零和游戏的平衡。非常流行的生成对抗网络提供了一个很好的例子，其中训练目标被框架为两个相互竞争的模块，用于搜索最小-最大解决方案。在使用学习系统寻找平衡方面已经做了大量的工作，研究者最近的工作表明，可以通过重复游戏中的学习来看待几个基本的凸优化过程。该奖项将有助于支持数学框架的进一步发展，将这些结果扩展到凸优化之外，并在非凸设置中设计具有可证明保证的高效算法。其中一个特别感兴趣的领域将是在训练复杂的多人问题中使用连续时间分析，以了解这种动态何时导致稳定的结果，何时引发混乱的行为。该奖项反映了美国国家科学基金会的法定使命，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，被认为值得支持。

Machine Learning systems are becoming ubiquitous and increasingly complex in modern life. Many devices such as mobile phones continually run dozens of predictive models, and these models receive input not only from the user but also from each other. One way to think about the unexpected challenges of multiple interacting learning systems is to consider how humans interact in personal relationships or even how governments engage with each other during international disputes. Such scenarios involve hard-to-predict dynamics, where the introduction of a small amount of information or minor changes to strategy can give rise to highly different outcomes. This project aims to understand these interacting dynamics from an algorithmic perspective, with an eye towards designing modular learning systems where the implementer can be certain that the dynamics of training will reach a desired solution. The work will significantly increase the range of tasks and challenges where learning systems are applied in the real world and will have a strong impact on how artificial intelligence interacts with society.The project begins with a focus on game theory and builds off of a number of both classical and recent results in solving so-called min-max problems, where one wants to find the equilibrium of a zero-sum game. The hugely popular Generative Adversarial Networks provide a great example where the training objective is framed as two competing modules engaged in a search for a min-max solution. There has been a great deal of work in finding equilibria using learning systems, and recent work by the investigator has shown that several fundamental convex optimization procedures can be viewed through the lens of learning in repeated play. The award will help support the further development of mathematical frameworks to extend these results beyond convex optimization and to design efficient algorithms with provable guarantees in non-convex settings. One of the areas of particular interest will be the use of continuous-time analysis in training complex multiplayer problems, to understand when such dynamics lead to stable outcomes and when they elicit chaotic behavior.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.