Editor's Note: From artificial intelligence Since it was first put forward, the controversy about it has never stopped. Now the time is ripe, and the market has begun to focus on it. Transforming Intel We also see that it has unlimited possibilities in the future. Intel hopes to become a leader in the industry by laying out all aspects of AI. This is the first media communication meeting held by Intel on AI in China.
(The editor has partially simplified the speech on Intel in this article without modifying the original meaning.)
This is an era of artificial intelligence
In fact, until today, many people still believe that AI is far away from us, but they do not know that Microsoft Cortana and iPhone are among the Windows devices Siri, recommendation engine, face recognition, image recognition, these technologies are based on the support of artificial intelligence. In fact, artificial intelligence has entered our lives. It is not as science fiction as described in novels and movies, but it is within reach.
The reason why AI has unlimited possibilities in the future is that AI involves too many fields. After hardware and software catch up, it will greatly change the way business operates and people live.
At the meeting, Intel believed that the era of artificial intelligence has come, which can be attributed to the following three reasons:
1、 At present, cloud computing has achieved large-scale expansion, and cloud computing is all over the world.
2、 With the improvement of hardware level, the cost of computing is constantly declining, and the economy of computing is becoming stronger and stronger.
3、 The number of connected devices is constantly increasing, the interconnection between devices is further improved, and the data has achieved explosive growth.
In fact, because AI needs a lot of high-performance computing, it has strict requirements on hardware and algorithms. The end of the PC era is precisely because performance computing has seriously overflowed in the personal PC field, while the field of artificial intelligence is just the beginning of high-performance computing, which is undoubtedly an exciting thing for Intel. The requirements of AI for high-performance computing are as endless as the requirements of games for GPUs. At least we still cannot see the performance overflow threshold of AI.
Machine learning in artificial intelligence
When we are paying attention to artificial intelligence, we are actually paying attention to machine learning. Machine learning is a branch of artificial intelligence, which is currently the fastest growing branch of artificial intelligence. Because AI is constantly learning machines to become more intelligent, here is a familiar example - Google's Alpha Go. This year's Go competition between AlphaGo and Li Shishi has attracted strong attention from all walks of life. Five months before the battle between AlphaGo and Li Shishi, AlphaGo defeated the European Go champion Fan Hui Er Duan. Until the end of last year, its Go level was 3168, while Li Shishi, the second in the world, had 3532 points. According to this level, AlphaGo's odds of winning each set were only about 11%. As a result, three months later, AlphaGo won 4-1 against Li Shishi. Its learning ability was so fast that people were worried.
Machine learning needs not only strong hardware support, but also further optimization of hardware, Intel is very proud to think that its chips (Xeon Phi, FPGA), storage (3DXpoint flash memory technology), sensors (including RealSense), libraries, and reference architectures can enable developers to better carry out research on machine learning, which is an all-round layout. Nvidia, another leader in the field of artificial intelligence, focuses on violently improving the GPU performance to accelerate the development of machine learning. Nvidia and Intel both hope to extend their leadership in the PC industry to the field of artificial intelligence, but their strategies for artificial intelligence are different.
Combination of artificial intelligence and machine
Dr. Song Jiqiang, president of Intel China Research Institute, said that the development process of intelligent machines was not achieved overnight. It went through many stages. In general, there are three stages:
1、 Interconnection. From the original embedded devices that cannot be networked to today's networked machines, they are no longer isolated.
2、 Intelligence. Through software and hardware, we can realize perception and processing, and allow machines to interact with humans in an advanced way. This era is basically an era that we have developed through smart phone technology.
3、 Autonomy. Autonomous machines need a deep technical understanding. At this stage, artificial intelligence can play a major role, including how machines plan, reason, predict, and finally make correct processing and feedback. In this process, the behavior of machines must be reliable.
We are now in the late stage of "intelligence". From intelligence to autonomy, Intel looks at AI from the perspective of computing and analyzes two challenges we are facing at present:
1. Data exchange between multi-sensor and real-time feedback of machine
2. Is AI on the device or the cloud?
For the first challenge. Due to the complexity of AI interaction, AI machines often use many sensors, which collect image signals, sound signals, energy signals, biological signals, electrical signals, etc. Of course, acquisition is not enough for them to process. Here, Intel takes an example. The robot that handles some unusual things in aerospace is called "Mantis Robot". There are 4000 sensors, many of which are visual. At this time, there are many data streams. It is a great challenge to process and feedback so many data streams while inputting.
For the second challenge, Intel also gave some examples. For example, for facial expression recognition, there are seven basic facial expressions that can be detected, which requires 100 frames/second of image processing speed. Because this interaction process requires a very fast speed, it is too slow to process through cloud gold and then feed back to the device side, so we can only do this at the device side. As for unmanned vehicles, the front end is mainly responsible for perception, and the data processing after perception is handed over to the cloud. Because unmanned vehicles involve a lot of information, not only simple image recognition, but also some environmental recognition and biometric recognition. Because the data processed is large and comprehensive, and must have security, the cloud is needed at this time. There is also a transmission rate and delay problem involved here. 5G network is very necessary.
If Intel's product layout for AI is simple, it can be divided into the following layers:
Xeon Phi+ Nervana : Used for high-performance computing at the top level of the cloud.
Xeon+FPGA : Used for low power performance calculation of cloud M-server/front-end devices.
Core（GT） : Used for performance calculation and graphics acceleration of consumer front-end devices.
Euclid : The development board provided to developers/makers integrates Atom low-power processor, RealSense camera module and interface, which can be used as the core development component of UAVs and small robots.
Curie : The module provided to developers/makers, with built-in Quark SE system chip, Bluetooth low-power radio, accelerometer, gyroscope and other sensors, can be used as the core component of low-power wearable devices.
With regard to 5G, Intel has been promoting the development of NFV (Network Function Virtualization). At present, China Mobile has begun to deploy NFV related hardware devices on a large scale, and Huawei and Intel are also partners in this regard.
From micro sensors to enterprise class cloud processors, from developers to enterprises, Intel hopes to infiltrate every link of AI through this overall and comprehensive end-to-end link, so as to form a virtuous circle of development.
