SBIR Phase I: A Hardware-Aware AutoML Platform for Resource-Constrained Devices

这个小企业创新研究（SBIR）第一阶段项目的广泛影响将是自动化制造检查，降低某些高度工程化项目的成本。该项目将开发一个系统，为资源受限的工业硬件自动创建机器学习模型。该项目能够快速、经济、轻松地创建具有重要工业应用的设备上人工智能。这将改进健壮的视觉检测工具，以提高产品质量。该项目将开发一种高效的解决方案，以创建先进的工业物联网应用，减少网络压力，最大限度地减少延迟，并提高边缘的安全性。虽然物联网在制造业中有许多应用，但由于其低延迟需求和它在配套基础设施中产生的高压力，视觉检测市场可以受益。该项目是一个灵活的模块化硬件感知机器学习模型生成系统，通过为资源受限的设备自动生成复杂的机器学习模型来减少人工工作量。技术障碍包括在没有人为干预的情况下考虑硬件速度和容量限制的神经网络的生成。技术里程碑包括创建针对边缘设备定制的自动模型发现平台、新的资源约束神经架构搜索算法和硬件感知模型压缩。这项研究旨在生产一个原型，供计算机科学培训有限的领域专家用于创建先进的工业物联网解决方案。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。

The broader impact of this Small Business Innovation Research (SBIR) Phase I project will be to automate manufacturing inspection, reducing the cost of certain highly engineered projects. This project will develop a system to automatically create machine learning models for resource-constrained industrial hardware. This project enables the fast, affordable, and easy creation of on-device artificial intelligence with significant industrial applications. This will improve robust vision-inspection tools to improve product quality. This project will develop an efficient solution to create advanced industrial internet of things applications that reduce network stress, minimize latency, and increase security at the edge. While the internet of things has many applications in the manufacturing industry, the vision inspection market can benefit due to its low-latency needs and the high stress it produces in the supporting infrastructure. This project is a flexible and modular hardware-aware machine learning model generation system that reduces manual efforts by automatically generating complex machine learning models for resource-constrained devices. Technical hurdles include the generation of neural networks that consider hardware speed and capacity constraints without human intervention. Technical milestones involve creating an automated model discovery platform tailored for edge devices, new resource-constrained neural architecture search algorithms, and hardware-aware model compression. This research aims to produce a prototype that domain experts with limited computer science training can use to create advanced industrial internet of things solutions.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.