NSF-BSF: AF: Small: An Algorithmic Theory of Brain Networks

理解大脑作为一种计算设备是如何工作的，是现代神经科学和人工智能的核心挑战。不同的研究团体以不同的方式应对这一挑战，包括将神经网络结构作为计算功能的线索，使用功能成像研究神经激活模式，基于神经计算的简化模型开发理论，以及设计神经启发的机器学习体系结构。这个项目将使用分布式计算理论和其他理论计算机科学分支的技术来解决这个问题。这个项目有可能通过确定大脑中解决的关键问题和可能用于解决这些问题的关键机制来提高我们对大脑计算的理解。这项工作也可以对理论计算机科学产生影响，为理论研究提供一个新的、富有成效的方向。麻省理工学院和以色列魏茨曼研究所的这项合作将增加女性和少数族裔参与者在这一领域的参与，并将寻求弥合计算机科学家和生物学研究人员之间的差距。具体地说，该项目开发了一种脑网络算法理论，基于具有一般互连模式的新型随机尖峰神经网络模型。它定义了这些网络要解决的抽象问题的集合，灵感来自于在实际大脑中解决的问题，如注意力、识别、学习和记忆问题。该项目设计了解决问题的算法(网络)，并根据静态成本(如网络规模)和动态成本(如收敛到正确解决方案的时间)对其进行分析。研究人员考虑了各种成本之间的权衡，并将证明相应的下界结果。模型、问题和解决方案应该足够简单，以便进行理论分析，但又足够现实，可以深入了解真实神经网络的行为。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Understanding how the brain works, as a computational device, is a central challenge of modern neuroscience and AI. Different research communities approach this challenge in different ways, including examining neural network structure as a clue to computational function, using functional imaging to study neural activation patterns, developing theory based on simplified models of neural computation, and engineering of neural-inspired machine learning architectures. This project will approach the problem using techniques from distributed computing theory and other branches of theoretical computer science. This project has the potential to improve our understanding of computation in the brain, by identifying key problems that are solved in the brain and key mechanisms that may be used to solve them. This work can also have impact on theoretical computer science, by contributing a new and fruitful direction for theoretical study. This collaboration between MIT and the Weizmann Institute in Israel will increase the participation of women and minority participants in this field and will seek to bridge the gap between computer scientists and biology researchers.Specifically, the project develops an algorithmic theory for brain networks, based on novel stochastic Spiking Neural Network models with general interconnection patterns. It defines a collection of abstract problems to be solved by these networks, inspired by problems that are solved in actual brains, such as problems of focus, recognition, learning, and memory. The project designs algorithms (networks) that solve the problems, and analyze them in terms of static costs such as network size, and dynamic costs such as time to converge to a correct solution. The investigators consider tradeoffs between the various costs, and will prove corresponding lower bound results. The models, problems, and solutions should be simple enough to enable theoretical analysis, yet realistic enough to provide insight into the behavior of real neural networks.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.