High performance computer vision on low performance hardware

人们普遍认为机器学习有助于数据分析和知识转化。**特别是，研究表明，机器学习驱动的监测和分析可以提高运动员的表现和技能(goo.gl/rtHfhk)。技术(特别是可穿戴技术)已经成为体育产业的关键组成部分。有几家公司专注于可穿戴技术，但**它们利用机器学习的能力受到这些可穿戴传感器中的计算硬件的限制；**通常是嵌入式微控制器和处理器。嵌入式处理器通常受到**内存和功耗的限制，这使得机器学习模型如何在**不损失任何精度的情况下可靠地部署尚不清楚。**Form Athletica开发可穿戴技术来帮助运动员收集有关他们活动的信息，并**向运动员提供反馈。目前，Form的可穿戴设备主要收集和测量活动数据，并**使用离线机器学习来确定对运动员的反馈。由于可穿戴设备上的内存太小，来自可穿戴设备的数据受到限制，给运动员的反馈很慢，而且**缺乏特异性。Form正在让他们的服务变得实时，并将机器学习转移到**可穿戴设备本身。我们的工作将寻求优化和压缩机器学习模型，以便在**可穿戴设备上运行，并利用传感器已经生成的数据。我们的主要目标是开发利用数据压缩、模型压缩、指令**布局的**系统优化和实现，以支持在嵌入式处理器上部署ML算法；以前的工作主要集中在**耗费资源的GPU上。除了测试其核心功能和实现，我们还将探索压缩机器学习算法向运动员提供的反馈的**准确性。

It has become widely accepted that machine learning can help with data analysis and conversion to knowledge.**In particular, studies have shown that machine-learning driven monitoring and analysis can increase athlete**performance and skills (goo.gl/rtHfhk). Technology (and specifically wearable technology) has become a**critical component of the sports industry. There are several companies that focus on wearable technology, but**their ability to leverage machine learning is limited by the computing hardware in these wearable sensors;**typically embedded micro-controllers and processors. Embedded processors are typically are limited by**memory and power consumption, making it unclear the how machine learning models can be deployed reliably**without losing any accuracy.**Form Athletica, develops wearable technology to help athletes collect information about their activity and**provide feedback to athletes. Currently, Form's wearable devices primarily collect and meter activity data, and**use offline machine learning to determine feedback to the athletes. The feedback to the athletes is slow and**lacks specificity due to the data from wearable devices being limited by the small amount of memory on the**device. Form is in the process of making their service real-time and moving the machine learning onto the**wearable device itself. Our work will seek to optimize and compress the machine learning models to run on the**wearable device and leverage the data already being generated by the sensors. Our primary goal is to develop**system optimizations and implementations that leverage data compression, model compression, instruction**layouts to enable the deployment of ML algorithms on embedded processors; prior work has largely focused on**resource hungry GPUs. Beyond testing its core functionality and implementation, we will also explore the**accuracy of the feedback provided by the compressed machine learning algorithms to athletes.