Earlier this year, Apple announced its US$200 million acquisition of Seattle-based edge-AI startup Xnor.ai. This was one of its many other moves to bring the AI-inferencing from cloud to the local hardware.

今年早些时候，苹果宣布以2亿美元收购总部位于西雅图的Edge-AI初创公司Xnor.ai。 这是将AI推理从云引入本地硬件的众多举措之一。

Today, algorithms sit on the cloud and serve your requests. As you can guess, there can be a latency as the data has to travel between the cloud servers and the remote client. There might also be many privacy concerns attached. The steady growth in the power of smart-devices is slowly giving rise to the opportunity to run the algorithms locally on edge instead of on the cloud. All major tech-giants are now in the race to push AI to the edge.

如今，算法位于云上，可以满足您的请求。 您可能会猜到，由于数据必须在云服务器和远程客户端之间传输，因此可能会有延迟。 可能还会涉及许多隐私问题。 智能设备功能的稳定增长正逐渐带来在本地而不是在云上运行算法的机会。 现在，所有主要的技术巨头都在争先恐后地推动AI的发展。

“Machine learning is going to happen at the edge in a big way, The big question is how do you do it efficiently?” says Stanford Professor Subhasish Mitra.

“机器学习将在很大程度上发生在边缘，最大的问题是您如何有效地做到这一点？” 斯坦福大学教授Subhasish Mitra说。

什么是Edge AI？ (What is Edge AI?)

In Edge computing, data is processed by the device itself or by a local computer or server, rather than being transmitted to a data center, says HP. Now when we talk about AI, these computations become quite complicated, and ordinary edge devices become incapable of performing them. That is why presently, these complex computations are left to the hefty servers at the data center.

惠普表示，在Edge计算中，数据由设备本身或由本地计算机或服务器处理，而不是传输到数据中心。 现在，当我们谈论AI时，这些计算变得非常复杂，普通的边缘设备也无法执行它们。 这就是为什么目前将这些复杂的计算留给数据中心的大型服务器的原因。

If you have ever used FaceApp (Application to try different looks), you might have noticed the considerable amount of time that it takes to give you the result. What’s happening on the backend is that your image(encrypted) is being sent to the remote server, the server applies the algorithm requested by you and sends it back to your device. Now, if that complete process somehow takes place in your local itself, it would be much faster and not mention much secure. This is where Edge-AI comes in, and companies like Xnor.ai are building tech that enables local inferencing possible.

如果您曾经使用过FaceApp(应用程序尝试不同的外观)，那么您可能已经注意到花费大量时间才能获得结果。 后端发生的事情是(加密的)图像已发送到远程服务器，服务器将应用您请求的算法并将其发送回设备。 现在，如果完整的过程以某种方式发生在您自己的本地本身中，它将更快并且不会提及太多安全性。 这就是Edge-AI的用武之地，而Xnor.ai之类的公司正在开发使本地推理成为可能的技术。

Edge AI为什么重要？ (Why does Edge AI matter?)

Edge AI gives rise to a possibility for mission-critical and time-sensitive decisions to be made faster, more reliable, and with greater security. For Apple and many other companies that claim to put the privacy of a user above all, edge AI would definitely be a go-to piece of technology.

Edge AI使得关键任务和对时间敏感的决策得以更快，更可靠且具有更高安全性的可能性成为可能。 对于苹果公司和许多声称将用户隐私放在首位的公司而言，边缘AI绝对是必不可少的技术。

Xnor.ai’s technology runs deep learning models efficiently ​on edge devices such as phones, IoT devices, cameras, drones, and embedded CPUs. The trick is to build highly efficient hardware that is capable of running power-hungry algorithms without much loss in the accuracy of these algorithms. Apple’s acquisition of Xnor.ai will potentially help it challenge the big guys in cloud technology, Amazon Web Services, Microsoft Azure, and Google Cloud. On the otherwise, all of these are also continuously increasing their focus on harnessing the benefits of AI and edge computing.

Xnor.ai的技术在电话，IoT设备，摄像头，无人机和嵌入式CPU等边缘设备上高效地运行深度学习模型。 诀窍是构建高效的硬件，该硬件能够运行耗电大的算法，而又不会大大降低这些算法的准确性。 苹果对Xnor.ai的收购将有可能帮助其挑战云技术，亚马逊网络服务，微软Azure和Google Cloud等大公司。 否则，所有这些都将继续增加他们对利用AI和边缘计算的优势的关注。

您如何优化优势 (How do you optimize for the edge)

There can be two broad ways in which you can go about improving the performance of the algorithms on edge: Optimizing the model architecture, Optimizing the device hardware, or both.

可以通过两种广泛的方法来提高边缘算法的性能：优化模型体系结构，优化设备硬件，或两者。

Optimizing the model architecture

优化模型架构

Traditional AI algorithms use a conventional representation of numbers. For example, your 32-bit PC might be using 32-bit weights and biases for the model. Over the years, what people have observed is that the weights and biases don’t need to be precise to 32 or 64-bit to give good accuracies. Studies show that when 32-bit precision weights and biases are replaced by 8 or 16-bit precision, there is no significant drop in the accuracy of the model.

传统的AI算法使用数字的常规表示形式。 例如，您的32位PC可能在模型中使用32位权重和偏差。 多年来，人们观察到，权重和偏差不必精确到32位或64位就可以提供良好的精度。 研究表明，当将32位精度权重和偏差替换为8位或16位精度时，模型的精度不会显着下降。

One Publication by Xnor.ai titled ImageNet Classification Using Binary Convolutional Neural Networks, published in 2016, proposes two efficient approximations to standard convolutional neural networks: Binary-Weight-Networks and XNOR-Networks. They reduce the precision to 1-bit(binary) in order to save memory by 32 times. The XNOR-Networks approximate convolutions using primarily binary operations resulting in 58 times faster computation. The final scores were excellent; the paper says, “the classification accuracy with a Binary-Weight-Network version of AlexNet is the same as the full-precision AlexNet,” which is purely amazing. One thing to note is that the paper was published in 2016 and, there are better architectures today.

Xnor.ai于2016年发布的一篇名为《 使用二进制卷积神经网络的ImageNet分类 》的出版物提出了对标准卷积神经网络的两种有效近似：Binary-Weight-Networks和XNOR-Networks。 它们将精度降低到1位(二进制)，以节省32倍的内存。 XNOR网络主要使用二进制运算来近似卷积，从而使计算速度提高了58倍。 最终分数非常好； 该论文说：“使用Binary-Weight-Network版本的AlexNet进行分类的准确性与使用完全精度的AlexNet相同，”这真是令人惊讶。 要注意的一件事是，该论文于2016年发布，如今有更好的架构。

Optimizing the Hardware

优化硬件

The next approach could be building better hardware to support compute and power-hungry AI algorithms. Myriad was one such company working on developing hardware units, especially for AI Inferencing. In 2016, AI acquired Movidius and the hardware today is known as Intel’s Neural Compute Stick. The Neural Compute Stick or NCS allows you to develop, fine-tune, and deploy traditional convolutional neural networks (CNNs) on low-power applications that require real-time inferencing.

下一种方法可能是构建更好的硬件来支持计算和耗电的AI算法。 Myriad就是这样的公司之一，致力于开发硬件单元，尤其是用于AI推理。 AI在2016年收购了Movidius，如今的硬件被称为英特尔的神经计算棒。 神经计算棒或NCS允许您在需要实时推理的低功率应用程序上开发，微调和部署传统的卷积神经网络(CNN)。

NCS Stick works with Intel’s OpenVINO toolkit that helps you to take your traditional machine learning model, generate Intermediate Representations, and then use the Inference Engine to infer the result for a given input data. The intermediate representations are generated by the Model Optimizer module of the toolkit that converts your model to a form that’s faster in inferencing.

NCS Stick与Intel的OpenVINO工具包一起使用，可帮助您采用传统的机器学习模型，生成中间表示，然后使用推理引擎来推断给定输入数据的结果。 中间表示由工具箱的“模型优化器”模块生成，该模块将模型转换为推理速度更快的形式。

The best method would, however, be the software-hardware co-design of the entire pipeline. This would ensure that your performance is optimized on both hardware and software levels. In the video below you, can see how Xnor.ai does a fantastic job of building a solar-powered object detection chip. The use of such chips can span thousands of real-life applications.

但是，最好的方法是整个管道的软件-硬件协同设计。 这样可以确保在硬件和软件级别上都优化了性能。 在下面的视频中，可以看到Xnor.ai如何出色地完成了构建太阳能物体检测芯片的工作。 这种芯片的使用可以跨越数千个现实生活中的应用。

对我们有什么帮助 (What’s in hold for us)

As Edge AI gets more powerful and starts taking shape, we will witness true-AI on our devices. Our smart devices would no longer be required to connect to the internet to access AI-based applications; the latency would reduce drastically, and above all, true-privacy would be ensured.

随着Edge AI的功能越来越强大并开始成形，我们将在设备上见证true-AI。 我们的智能设备将不再需要连接到Internet即可访问基于AI的应用程序； 延迟将大大减少，最重要的是，将确保真正的隐私。

It’s one exciting time to be alive with a lot of exciting stuff happening across the globe. One thing that I definitely look forward to is how edge AI influences healthcare once it is mature. Thanks for reading the article and I hope you enjoyed it. See you soon :)

这是活着令人兴奋的时刻，全球各地发生了许多令人兴奋的事情。 我绝对期待的一件事是，边缘人工智能一旦成熟，将如何影响医疗保健。 感谢您阅读本文，希望您喜欢它。 再见 ：)