Collaborative Research: NCS-FR: DEJA-VU: Design of Joint 3D Solid-State Learning Machines for Various Cognitive Use-Cases

现代计算机使需要快速大规模算术运算的广泛应用发生了革命性变化。然而，今天的计算机在“认知”任务上的表现仍然很差--这些任务需要建立知识，从过去的经验中学习，并预测未来。该项目旨在通过利用最近对大脑如何表示和计算信息的理解的进步，以及将大脑结构特征的复杂信号路径映射到三维集成芯片上的关键见解，设计一种新型计算机芯片--认知计算机器(C2M)。该项目的灵感来自海马体，这是一种已知对执行这些认知任务至关重要的大脑结构。该项目将对海马体中的关键信息处理步骤进行建模和量化。然后，这些关键的海马体功能将通过最先进的硬件设计技术嵌入到固态计算芯片上。海马区感知、硬件感知的算法框架将增强芯片设计努力，使其能够在资源受限的环境中进行在线学习和决策。该项目在工业、消费和国防领域的机器人和自主系统领域具有潜在的颠覆性应用。每个参与的研究人员和机构都致力于支持广泛的培训和指导计划，重点是来自科学界代表性不足的群体的学生。参与该项目的学员将接受罕见的神经科学和工程学跨学科培训，为各种职业发展轨迹奠定基础。此外，参与的实验室将通过科学文章、公开和会议演示以及其他外联工具传播项目成果，包括项目网站和联合课程活动。利用个别事件的信息建立知识并做出适合背景的决定的认知能力是人类日常生活不可或缺的，但对硬件和软件系统构成重大挑战。数十年的研究表明，一种名为海马体的大脑结构在做出与环境相适应的决策方面起着至关重要的作用。这个项目的目标是设计一类新的计算机芯片(认知计算机器或C2M)，灵感来自海马体的认知功能。该项目利用了三项重大和及时的发展：(1)芯片的三维集成，实现了时空信号的新型路由技术；(2)记忆中处理技术，能够进行复杂的动态模拟芯片处理；以及(3)在理解海马体机制和支持学习、记忆和决策的动力学方面取得了进展。此外，在Silico C2M中设计的将通过适用于一系列认知应用的新颖的海马区感知、硬件感知的学习算法，增强快速和稳健的决策能力。该项目的变革潜力来自于在三个不同抽象水平(线程)进行的研究，并指向一个共同的目标：(1)神经科学抽象，如确定和回答有关海马体组织和功能原理的关键问题；(2)硬件抽象，如以技术友好的方式在功能上模拟3D集成电路中的海马体计算属性；以及(3)算法抽象，例如将来自受海马体启发的芯片的基于事件的预测与基于知识的预测相结合，以实现快速和稳健的学习。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。

Modern computers have revolutionized a wide range of applications that require fast, large-scale arithmetic operations. However, today’s computers still perform poorly on ‘cognitive’ tasks – tasks that require knowledge building, learning from past experiences, and anticipating the future. This project aims to design a new class of computer chips – ‘Cognitive Computing Machines (C2M)’ – by leveraging advances in recent understanding of how the brain represents and computes information and the crucial insight to map complex signal routing, characteristic of brain structures, onto three-dimensional integrated chips. The inspiration for the project is the hippocampus, a brain structure known to be critical for performing these cognitive tasks. The project will model and quantify key information processing steps in the hippocampus. These key hippocampal functions will then be embedded on to solid-state computing chips through state-of-the-art hardware design techniques. A hippocampal-aware, hardware-aware, algorithmic framework will augment the chip design efforts to enable online learning and decision-making in resource constrained environments. The project has potential disruptive applications in the field of robotics and autonomous systems spanning industrial, consumer and defense sectors. Each participating investigator and institution are committed to support a wide range of training and mentoring programs, with a focus on students from groups underrepresented in science. Trainees involved in the project will receive rare cross-disciplinary training in neuroscience and engineering, providing a foundation for a wide variety of career trajectories. Further, the participating laboratories will disseminate project outcomes through scientific articles, public and conference presentations, and other outreach tools including project websites and joint curricular activities.The cognitive ability to use information from individual events to build knowledge and make context-appropriate decisions is integral to daily life in humans but poses a significant challenge for hardware and software systems. Decades of research has indicated that a brain structure called the hippocampus plays a crucial role in enabling context-appropriate decision-making. The goal of this project is to design a new class of computer chips (Cognitive Computing Machines or C2M) inspired from the cognitive functions of the hippocampus. The project leverages three significant and timely developments: (1) three-dimensional integration of chips, enabling novel routing techniques for spatio-temporal signals, (2) processing-in-memory technology, capable of complex dynamic analog on-chip processing, and (3) advances in the understanding of hippocampal mechanisms and dynamics supporting learning, memory, and decisions. Further, the designed in silico C2M will be augmented with rapid and robust decision-making through novel hippocampus-aware, hardware-aware learning algorithm for range of cognitive applications. The transformative potential of the project emerges from research conducted at three different levels of abstractions (threads) and directed towards a common goal: (1) neuroscience abstraction, as in identifying and answering key questions about organizational and functional principles of hippocampus; (2) hardware abstraction, as in functionally mimicking hippocampal computing attributes in 3D integrated circuits in a technology-friendly manner; and (3) algorithm abstraction, as in incorporating event-based predictions from the hippocampus-inspired chip with knowledge-based predictions for rapid and robust learning.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.