I-Corps: Algorithm-Hardware Co-Design for Large-Scale Machine Learning

这个i-Corps项目的更广泛的影响/商业潜力是减少了阻碍采用机器学习技术当前和未来进步的经济和技术障碍。人工智能(AI)的进步推动了机器学习模型在计算上变得越来越大。由于培训和部署这些模型所需的计算和内存要求，目前能够使用最复杂模型的组织相对较少。这种缺乏访问权的情况导致试图将人工智能应用于机器人、自然语言处理、药物发现和计算机视觉等应用的研究人员和企业的成本很高。这里开发的技术缩小了模型的尺寸并优化了硬件，以便大型模型可以在现有系统上使用。改进对最先进模型的访问将加快人工智能在经济中的采用，并潜在地推动基于人工智能的尖端发现的更快改进。这个i-Corps项目开发了机器学习加速领域的几项创新的组合。这项创新旨在硬件和软件层面的同步优化。通过同时瞄准这两个级别，该技术实现了通过单独结合现有技术无法实现的速度。核心技术结合了模型压缩技术、用于加速的定制内核和硬件利用设计以及云编排算法。先前的结果表明有能力显著减少模型大小和加速推理结果。这一奖项反映了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。

The broader impact/commercial potential of this I-Corps project is the reduction of the economic and technical barriers preventing the adoption of current and future advancements in machine learning technologies. Advances in artificial intelligence (AI) have pushed machine learning models to be computationally larger and larger. Relatively few organizations currently have access to the most sophisticated models due to the necessary computational and memory requirements required to train and deploy such models. This lack of access results in high costs for researchers and businesses attempting to apply AI in applications such as robotics, natural language processing, drug discovery, and computer vision. The technology developed here reduces the size of the models and optimizes hardware so that large-scale models can be used on existing systems. Improved access to state-of-the-art models will accelerate adoption of AI in the economy and potentially drive more rapid improvements in cutting edge AI-based discoveries.This I-Corps project develops a combination of several innovations in the field of machine learning acceleration. The innovation is aimed at simultaneous optimization at the hardware and software levels. By targeting both levels simultaneously, the technology enables speeds which are not possible by combining existing technologies individually. The core technology combines model compression techniques, custom kernels and hardware utilization designs for acceleration, and cloud orchestration algorithms. Prior results have shown capabilities to significantly reduce model sizes and speed inference results.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.