CAREER: Efficient Algorithms and Hardware for Accelerated Machine Learning

人工智能（AI）正在为人类社会带来巨大的利益，在解决人类未来的基本问题方面具有突出的潜力。深度神经网络在许多人工智能应用中显示出显著的进步。然而，这种好处是以高昂的计算资源和工程师资源为代价的。对于功率预算紧张的移动设备，如此高的计算要求变得令人望而却步。另一方面，设计神经网络的专家也很缺乏，众所周知，神经网络很难调整。本项目系统地研究高效的神经网络架构及其硬件加速器，使其在低功耗下快速运行。它还旨在通过基于AI的设计自动化（AI设计的AI）加快设计周期。这样的设计方法可以支持机器学习的研究和教育，同时显著提高机器学习模型的生产率。人们不再需要手动调整模型，但它是自动化的，使非专家能够构建高效的机器学习模型。它将使人工智能民主化，成为一个更多样化的社区。该项目旨在通过自动机器学习（AutoML）技术自动生成高效的神经网络及其硬件实现，这些网络可以推广到高维表示。它将设计一个硬件加速器，以提供更多的单位成本计算。算法硬件协同设计方法揭示了传统智慧所局限的更大的设计空间。它有望将神经结构搜索的设计周期比现有工作缩短两个数量级。借助AutoML、强大的硬件和共同设计的高效算法，可以在高维数据上解决更具挑战性的人工智能任务，这些任务以前很难或不可能被计算资源所阻碍，例如视频和3D点云。这些技术使人们对高维表示有了更深入的理解，并产生了最先进的神经网络架构，可以有效地在移动设备上运行，并保护用户隐私。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。

Artificial Intelligence (AI) is bringing significant benefits to human society with outstanding potential to address essential human concerns in the future. Deep neural networks have shown significant improvements in many AI applications. However, the benefit comes at the high cost of computational resources and engineer resources. For mobile devices with a tight power budget, such high demands for computation become prohibitive. On the other side, there is a shortage of experts who can design neural networks, which are notoriously hard to tune. This project systematically investigates efficient neural network architectures and their hardware accelerators to make them run fast at low power. It also aims to accelerate the design cycle by AI-based design automation (AI-designed AI). Such design methodology can support machine learning research and education,, while significantly improving the productivity of machine learning models. People no longer have to hand-tune the model, but it is automated, enabling non-experts to build efficient machine learning models. It will democratize AI to a more diverse community. The project aims to auto-generate both efficient neural networks and their hardware implementations that can generalize to high-dimensional representations, through automatic machine learning (AutoML) techniques. It will design a hardware accelerator to provide more computation per unit cost. The algorithm hardware co-design approach unveils a larger design space that conventional wisdom has been limited to. It is expected to shorten the design cycle of neural architecture search by two orders of magnitude over existing work. With AutoML, powerful hardware, and the co-designed efficient algorithms, it is possible to solve more challenging AI tasks on high-dimensional data that are previously difficult or impossible, hindered by the computation resource, such as videos and 3D point clouds. These techniques give rise to a deeper understanding of the high dimensional representations and produce state-of-the-art neural network architectures that can efficiently run on mobile devices as well as protect user privacy.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.