本文是TechXchange:AI在边缘
What you’ll learn
- 谁在开发机器学习解决方案？
- What types of data are being analyzed?
- What types of deployment platforms are popular?
机器学习（ML）是一个热门话题，但是许多人仍在学习曲线中，或评估ML是否适用于其应用程序。我们最近对电子设计读者，发现大多数受访者还发现ML不适用于其申请(Fig. 1)。The number of people using artificial-intelligence (AI) support to enhance their products or make them practical is small but growing.
某些应用程序使更广泛的采用due to volume, such as natural language processing in the cloud with platforms like Amazon Alexa and Apple’s Siri, as well as in applications like automotive advanced driver-assistance systems (ADAS). The payoff is significant and providing accelerated ML platforms is part of the design.
In the case of the cloud, processing audio and video data in parallel makes excellent use of cloud resources. Utilizing platforms such as smart NICs to service cloud computing make it possible for some ML models to reside in FPGAs in these adapters.
What I found most interesting was that most applications using ML employ multiple sensors(Fig. 2)。再说一次，这不应该太惊讶，因为复杂数据集的相关性和分析是深度神经网络（DNNS）在经过适当训练时可以做得很好的领域。
The emphasis on data streams like audio, video, and text/language processing should not come as a surprise either. These were some of the first areas to be addressed in research—numerous ML models can be trained and customized for a particular application. Building a new model is a much more difficult task, but one that can pay off in the long run.
基于软件的AI
One result I found very interesting was the dominance of software-based AI solutions(Fig. 3)。All platforms employ software somewhere to run the ML models. However, this question asked about the platform on which the models were run. It also was a multiple option question; the software could be running on ML-enhanced processors as well.
Still, the fact that software-based solutions are being employed without hardware acceleration indicates many applications don’t require additional hardware. We do know that many useful models can run on standard microcontrollers. The trick is matching the software to the application and running it on a hardware platform that will provide suitable performance to make the application work. This isn’t any different from the challenge faced by engineers and programmers when working on any project, since unlimited memory, computational power, and communication isn’t an option.
Another interesting piece of data is the low percentage of developers who rely on the cloud. This implies that most are involved with standalone ML applications. It also means that the hardware and software AI options available to developers are sufficient to incorporate ML models into their applications.
工具类型
Finally, we asked about the tools developers were using(Fig. 4)。I probably should have included a question about their use of development kits like those we’ve been highlighting in ourKit Close-Up video series。
Suffice it to say, every major chip vendor has an array of kits, reference designs, and tools that cater to AI applications. The only difference between them is the depth and breadth of the support, and the number of app notes and applications included with these kits. In general, it’s substantial and growing. It mirrors the rise of software support for processing hardware in general over the past couple decades, but it has taken a lot less time.
此过程背后的推动力归功于开源软件。工具的可用性突出显示了如何共享它们在采用方面有很大的不同。供应商提供的一系列开源软件也很重要。封闭式软件的量较小，但通常提供两种类型的供应商，并应用了封闭式支持来制作“秘密酱”。如今，由于模型是应用程序成功的关键，因此可以解决很多酱汁。
不管AI/ML的猖ramp，该技术仍处于起步阶段。并不是说这还没有准备好黄金时段 - 变化和改进的量是很大的。随着时间的推移，工具，平台和应用程序之间的差异很大，并且随着改进和新方法的添加，它会继续进行。
同样，对AI/ML可以发挥作用的理解仍在增长。与许多工具（例如用于图形的射线跟踪）不同，使用AI/ML技术为特定应用程序使用的适用性，适用性和实用性并不一定是显而易见的。
AI/ML isn’t a magic bullet and it can’t address all aspects of computing. We still need to code in C or other programming languages for the bulk of most applications. Nonetheless, AI/ML is simplifying many aspects of these applications, or it may make new features practical.
阅读更多文章TechXchange:AI在边缘
