CAREER: Designing Ultra-Energy-Efficient Intelligent Hardware with On-Chip Learning, Attention, and Inference

构建能够像人类一样执行认知任务（如学习、识别）的智能计算机一直是计算研究的长期目标。最近，最先进的深度学习和神经形态算法提高了认知应用软件的性能。然而，这种算法是计算-内存-通信密集型的，这使得硬件设计在便携式平台上执行低功耗实时训练和分类具有挑战性。此外，为了优化系统级功率，大规模硬件系统的高效供电和电源电压调节也成为一个关键问题。该项目将跨越硬件和软件设计的多个学科来解决这些挑战，朝着构建大脑启发的智能计算系统的总体目标迈进，这些系统在计算机视觉、语音、机器人和生物医学应用中的各种认知任务中都是超节能的。这项研究的成功可能会影响社会和工业中许多以用户为中心的计算系统，包括可穿戴、移动和边缘计算。该项目还包括综合教育和推广计划，包括新的跨学科课程开发、本科生/研究生培训和高中生暑期推广计划。在该项目中，将设计节能电路、架构和算法，以在面积/功率受限的移动/可穿戴硬件平台中整合学习、注意力和推理计算。为大幅度提高能源效率而将开发的具体技术包括：(1)基于生物启发注意力模型的最先进深度学习算法的计算冗余最小化；(2)适用于软件和硬件实现的新颖内存压缩方案；(3)在移动/可穿戴设备上消耗低功耗的实时片上学习方法；(4)能够适应认知工作负载突变的高效片上稳压器；(5)电路、架构和算法的跨层优化。这项研究的成果将以新的超大规模集成（VLSI）系统为特色，该系统可以以卓越的功率效率实时学习和执行认知任务，为小尺寸设备的无处不在的智能开辟了可能性。

Building intelligent computers that can perform cognitive tasks (e.g., learning, recognition) as well as humans do has been a long-standing goal of computing research. State-of-the-art deep learning and neuromorphic algorithms have recently advanced the software performance for cognitive applications. However, such algorithms are computation-memory-communication intensive, which makes the hardware design challenging to perform low-power real-time training and classification on portable platforms. Furthermore, to optimize system-level power, efficient power delivery and supply voltage regulation of such large-scale hardware systems also becomes a critical concern. This project will address these challenges across multiple disciplines of hardware and software design, towards the overarching goal of building brain-inspired intelligent computing systems that are ultra-energy-efficient for various cognitive tasks in computer vision, speech, robotics and biomedical applications. The success of this research is likely to impact many user-centric computing systems in society and industry, including wearable, mobile, and edge computing. This project also entails integrative education and outreach plans through a new interdisciplinary coursework development, undergraduate/graduate student training, and a summer outreach program for high school students.In this project, energy-efficient circuits, architectures and algorithms will be designed to incorporate learning, attention and inference computations in area-/power-constrained mobile/wearable hardware platforms. The particular technologies that will be developed to achieve large improvement in energy-efficiency include: (1) computation redundancy minimization of state-of-the-art deep learning algorithms with bio-inspired attention models, (2) novel memory compression schemes that apply to both software and hardware implementation, (3) real-time on-chip learning methods that consume low power on mobile/wearable devices, (4) efficient on-chip voltage regulators that can adapt to abrupt changes in cognitive workloads, and (5) cross-layer optimization of circuit, architecture and algorithm. The outcomes of this research will feature new very-large-scale integration (VLSI) systems that can learn and perform cognitive tasks in real-time with superior power efficiency, opening up possibilities for ubiquitous intelligence in small-form-factor devices.