Collaborative Research: SHF: Medium: Memory-efficient Algorithm and Hardware Co-Design for Spike-based Edge Computing

协作研究：SHF：中：基于 Spike 的边缘计算的内存高效算法和硬件协同设计

• 批准号: 2403723
• 负责人: Jae-sun Seo
• 金额: $ 50万
• 依托单位: Cornell University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-10-01 至 2027-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2403723&HistoricalAwards=false
• 关键词: Collaborative; Research; SHF; Medium; Memory

In today's rapidly advancing world of Artificial Intelligence (AI), energy efficiency has emerged as a crucial factor to facilitate the ubiquitous development of intelligent systems. The efficient deployment of AI holds the key to overcoming limitations posed by power-constrained devices and contributes to sustainable technological progress. Neuromorphic computing offers a brain-inspired paradigm of AI, called Spiking Neural Networks (SNNs), that represents a promising step forward in sustainable AI development. Inspired by the brain's neural architecture, SNNs process information in sparse, asynchronous, and event-driven patterns, resulting in reduced power consumption. This project aims to integrate SNNs with modern integrated circuits propelling energy efficiency across various AI domains, such as object detection, autonomous driving and image classification. The project team aims to devise novel algorithms and hardware design with prototype chips to accelerate the performance of SNNs in low-power and memory-efficient systems. These spiking neural chips will enable the practical and immediate application of neuromorphic systems in areas like drones, autonomous robots, portable medical devices, and wearable smart assistants. Furthermore, the project embraces an algorithm-to-system approach, providing opportunities for high school, undergraduate, and graduate students to explore research in the field of neuromorphic computing. An essential focus of this project also lies in training the next generation of scientists and engineers, fostering diversity, and promoting inclusivity within the AI and semiconductor fields. This project tackles the crucial task of enabling deep learning and AI algorithms on edge computing devices that have strict memory and power constraints. The key innovation lies in leveraging a brain-inspired spiking neural network (SNN) approach for edge computing. The team addresses the memory overhead issue of spiking neurons and takes a foundational approach, optimizing algorithms and hardware design for SNN deployment on edge devices. The project proposes algorithmic solutions, including novel architectures with shared computations and compression strategies, such as quantization and early exit. These optimizations aim to enhance the efficiency of SNNs on resource-constrained edge devices. On the hardware front, the project plans to demonstrate these ideas through prototype chip tapeouts with SNN-specific dataflow, event-addressable computations, and configurable support for proposed algorithm features. The goal is to develop a comprehensive understanding of the power, performance, and accuracy tradeoffs of SNNs for edge computing applications that will pave the way for sustainable AI.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.

在当今快速发展的人工智能（AI）世界中，能源效率已成为促进智能系统普遍发展的关键因素。人工智能的有效部署是克服功率受限设备所带来的限制的关键，并有助于可持续的技术进步。神经形态计算提供了一种受大脑启发的人工智能范例，称为尖峰神经网络（SNN），这代表了可持续人工智能发展的一个有希望的进步。受大脑神经结构的启发，SNN以稀疏、异步和事件驱动的模式处理信息，从而降低功耗。该项目旨在将SNN与现代集成电路集成，从而在各种人工智能领域（如物体检测、自动驾驶和图像分类）提高能效。该项目团队的目标是利用原型芯片设计新颖的算法和硬件设计，以加速SNN在低功耗和内存高效系统中的性能。这些尖峰神经芯片将使神经形态系统在无人机、自主机器人、便携式医疗设备和可穿戴智能助理等领域的实际和即时应用成为可能。此外，该项目还采用了算法到系统的方法，为高中生、本科生和研究生提供了探索神经形态计算领域研究的机会。该项目的重点还在于培养下一代科学家和工程师，促进多样性，并促进人工智能和半导体领域的包容性。该项目解决了在具有严格内存和功耗限制的边缘计算设备上实现深度学习和AI算法的关键任务。关键的创新在于利用大脑启发的尖峰神经网络（SNN）方法进行边缘计算。该团队解决了尖峰神经元的内存开销问题，并采取了一种基础方法，优化了边缘设备上SNN部署的算法和硬件设计。该项目提出了算法解决方案，包括具有共享计算和压缩策略的新架构，例如量化和提前退出。这些优化旨在提高资源受限边缘设备上SNN的效率。在硬件方面，该项目计划通过原型芯片流片来展示这些想法，这些原型芯片具有SNN特定的可编程逻辑、事件可寻址计算以及对拟议算法功能的可配置支持。该奖项旨在全面了解SNN在边缘计算应用中的功率、性能和准确性权衡，为可持续人工智能铺平道路。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，被认为值得支持。