SBIR Phase I: Enabling Real-Time AI on End Devices through Compression-Compilation Co-Design

这个小企业创新研究（SBIR）第一阶段项目的广泛影响在于它为扩大机器智能的使用创造了一系列新的机会。通过提供一种有效的方法来转换深度学习模型，以最佳地适应终端设备和实时应用程序的限制，该项目将大大缩短人工智能应用程序的上市时间，从而大大加快智能软件在健康、商业、金融、国防、社交网络和许多其他领域的开发和部署。这个小企业创新研究（SBIR）第一阶段项目旨在解决在终端设备（智能手机、无人机等）上有效开发实时人工智能应用的重要障碍。它通过一项突破性的技术——压缩编译协同设计来实现这一点。压缩和编译是在硬件上拟合深度学习模型以实现高效执行的两个关键步骤。模型压缩减小了深度学习模型的大小；编译从给定的深度学习模型生成可执行代码。压缩-编译协同设计的原则是将AI的两个组件以手拉手的方式进行设计。该技术采用一种新颖的方法，将一组新颖的模型压缩方法与压缩感知代码编译技术协同起来。其结果是，该技术实现了比目前技术水平高几倍的人工智能模型压缩率，几倍的速度，更好的能源效率和令人满意的准确性。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 is in the array of new opportunities it creates for expanding uses of machine intelligence. By providing an efficient way to transform deep learning models to best fit the constraints on end devices and real-time applications, this project will shorten the time to market for artificial intelligence applications by orders of magnitude, and hence significantly accelerate the development and deployment of intelligent software in health, commerce, financial, defense, social networks, and many other areas.This Small Business Innovation Research (SBIR) Phase I project aims to address the important barriers for efficient development of real-time Artificial Intelligence applications on end devices (smartphones, drones, etc.). It does this through a breakthrough technology, compression-compilation co-design. Compression and compilation are the two key steps in fitting a deep learning model on a hardware for efficient execution. Model compression reduces the size of deep learning models; compilation generates executable codes from a given deep learning model. The principle of compression-compilation co-design is to design the two components for AI in a hand-in-hand manner. The technology uses a novel approach to synergize a set of novel model compression methods with compression-aware code compilation techniques. The result is a technology that achieves several-fold higher artificial intelligence model compression rates over the state of the art, several times faster speed, better energy efficiency, and satisfying accuracy.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.