[May 31] news has been the application of artificial intelligence that there is a huge obstacle: the application of artificial intelligence algorithm runs very jumbled, and too dependent on high-power computer operator cloud computing and data centers, which makes it smart mobile phones and other "edge" applications on the device is not practical.
Now, however, thanks to a breakthrough in recent months, software, hardware, and a range of energy technologies, such concerns are rapidly dissipating, the new technologies are rapidly entering the market.
These new technologies may help artificial intelligence-driven products and services further away from dependence on powerful cloud computing services to enable them to enter our lives, we become even a small part of the body. Conversely, as the artificial intelligence services become ubiquitous, Deloitte (Deloitte) consulting firm late last year called "universal intelligence" may soon realize, it will shake up the industry in the coming years.
Market research firm Gartner Inc.'s data show that as of 2022, 80% of new smart phones shipped will have an independent artificial intelligence computing capabilities, this proportion is 10 percent higher than in 2017. ABI Research data show, is expected to 2023, shipments of smart phones will be about artificial intelligence computing capabilities of 1.2 billion, a figure much higher than the 79 million in 2017.
In this regard, a number of start-ups and their supporters have smelled a huge opportunity. Last week, the Embedded Vision Alliance (Embedded Vision Alliance) held a meeting in Silicon Valley. According to the alliance founder Jeff Bier said in the past three years, investors to set up new AI chip start-up companies invested a total of about $ 1.5 billion, an unprecedented scale. Market research firm Yole Developpment predicts that by 2023, a CAGR of artificial intelligence application processor from the current less than 20% to 46%, when almost all smart phones will be equipped with artificial intelligence processor.
Not only start-ups, established giants also join the competition team.
Just today, Intel demonstrated its upcoming glacial lake (Ice Lake) chip, which encompasses other instructions on the new artificial intelligence software and graphics processing unit "deep learning support" (Deep Learning Boost). On Monday morning, Arm also launched a series of processors for artificial intelligence applications, including smart phones and other high-end processor devices. After a few hours, NVIDIA released its first platform for artificial intelligence devices.
Tom Hackett IHS principal analyst at Markit, said: "Over the next two years, almost every processor vendors will offer some kind of artificial intelligence platform with competitive we are seeing now is a new opportunity.."
These chips is to enter the field more than smartphones. They are also used millions of "things" equipment, such as robotics, unmanned aerial vehicles, cars, cameras and wearable devices. About 75 development machine learning chip companies, to Israel's Hailo's case, the company had $ 21 million round of financing in January this year. In mid-May, the company released a processor designed for deep learning. Deep learning is a branch of machine learning, recently made a breakthrough in voice and image recognition.
New research shows that the scale neural network can be reduced 10-fold, at the same time, it can still achieve similar results. Thus, more compact and powerful software will pave the way for the end edge of artificial intelligence. Now, some companies are already trying to scale compression software needed for artificial intelligence.
For example, Google launched in late 2017 TensorFlow Lite machine learning library for mobile devices, which makes smart camera can identify the wildlife in the absence of an Internet connection, or offline help imaging equipment for medical diagnosis. Google staff research engineer Pete Warden keynote address at the Embedded Vision Summit said that currently there are about 2 billion mobile devices installed in the machine learning library.
In March this year, Google launched a speech recognizer, to provide support for the virtual keyboard, voice input applications Gboard for Google. The transcription speech recognition algorithm is less than 80 megabytes, so it can run offline on Arm A-series chips, and do not have network latency.
At the same time, concerns about privacy in the cloud rising rapidly, which means regulators have reason to avoid the exchange of data between sensitive equipment too often.
“几乎所有的机器学习运算都将在设备端完成，”Bier介绍道。Bier也是伯克利设计技术公司的联合创始人和总裁，该公司为嵌入式数字信号处理技术提供分析服务。这些设备的数目是巨大的。Warden指出，当今世界上约有2500亿个活跃的嵌入式设备，并且这个数字还在以每年20%的速度增长。
但在这样的设备上进行人工智能运算并非易事，这不仅取决于机器学习算法的大小，算法所需要的电量也格外重要。尤其是对于像智能手机、摄像头和传感器等物联网设备来说。它们不能过于依赖插座甚至电池供电。
云端收发数据的通讯过程能耗巨大，因此通过蜂窝网络或其他连接进行通信对许多小型廉价设备来说都是一个致命的障碍。Yole development的技术和市场分析师Yohann Tschudi表示:“我们需要一个专门的架构来完成低能耗的通讯。”
为此，还需要开发出一种实际能耗小于1毫瓦的设备，而这大约是智能手机用电量的千分之一。好消息是，越来越多的传感器甚至微处理器有望做到这一点。
例如，美国能源部与Molex 公司、SkyCentrics建筑公司合作，开发了用于建筑能源管理的低成本无线传感器。这个还在实验中的新型图像传感器可以利用环境光为自身供电。
而且，即使是计算的主力军——微处理器，也可以是非常低功耗的。 Warden说:“理论上讲，我们没有理由不能在微瓦或比毫瓦小一千倍的能耗下进行计算。”这在一定程度上是因为，它们可以被编程，例如，只有当一些特定的事情发生时，比如液体溅到地板上，它们才会唤醒与云端的数据交流。
所有这些都表明，机器学习在智能手机、智能相机和工厂监控传感器等领域有着广泛的新应用。Warden说道:“确实，我们收到了很多在嵌入式设备上运行机器学习的产品请求。”
这些应用包括:
使用加速度计预测维修，以确定机器是否震动太多或发出噪音。
街灯对行人的自动检测，只有当有人在附近时，街灯才会打开。
利用分散在农田中的视觉传感器或微型摄像机进行农业害虫识别。
Android phones using the old solar installation detection chain saw in a tree sound, to combat illegal logging.
When Warden even predicted that the sensor can communicate with each other, such as in smart home, smoke alarms to detect potential fire, toaster replied: toast No, it was just charred. These are only temporary speculation, but Google has not provided data on the use of intensive training, study "associative learning (federated learning)" to train a machine learning model (below).
Zhengzhou painless Hospital http://m.zyfuke.com/
Zhengzhou painless How much money http://mobile.zhongyuan120.com/