Song Jiqiang, President of Intel China Research Institute: 5G plays a key role in landing AI

Some experts in the communications industry pointed out that just as 3G opened the mobile Internet in 2007, commercial use of 5G will start the era of mobile AI in 2020. "When AI is deployed in a large scale in the application scenario, 5G has a crucial role in the landing of AI," said Song Jiqiang, director of the Intel China Research Institute, in an exclusive interview with the China Electronics News reporter. "We think AI's Applications are usually end-to-end scenarios. Data collection is at the front end. Data processing and value added are in the cloud. Value-added results must be returned to the front-end to improve the processing capabilities and processing performance of front-end devices. In closed-loop applications, there will be a lot of demand for 5G large-bandwidth, low-latency, and high-reliability networks. He believes that 5G also provides the possibility of using edges throughout the AI ​​architecture, allowing a lot of data to be quickly processed and recognized near the application scene.
AI applications rely on cost-effective networking
“We know that the rise of this wave of AI is basically based on deep learning. Depth learning requires a lot of data, and many data collection uses visual sensors such as cameras.” Song Jiqiang said, “If only AI training processes are considered This data can exist as long as there is a cloud, but in practice AI training needs to be performed in real time and constantly, in order to improve adaptability in different areas. For example, AI applications in unmanned or smart cities need such training.
According to Song Jiqiang, AI's reliance on the Internet is getting higher and higher, mainly reflected in the uploading of front-end collected data, data identification and analysis near the front-end, and the data value-added results are transmitted back to the front-end three links.
At the front end of AI's data collection, more and more wireless cameras are flooding and exploding. The development of smart homes and smart retailing brings more and more wireless cameras, which will also be needed in unmanned scenarios. Using multiple wireless cameras, these video data must be transmitted to the back end and transmitted to the cloud for data processing using AI. This process requires network support, especially for 5G networks with high capacity, large bandwidth, and low latency. Will be more suitable for AI data transmission needs.
“You can use the 5G feature to do training in the cloud and identify at the front end.” Song Jiqiang said, “The large amount of data needed for training can be processed at the front end, and if the data processing is placed on the front of the vehicle in the driverless scene. The processing will make the front end 'heavy', we can put the data processing at the edge of the network, near the front end, and the edge calculation is just 5G to bring."
"If there is no such network as 5G, artificial intelligence equipment not only does data collection in the front-end, but also analyzes and trains large amounts of data in the front-end, which makes artificial intelligence equipment very expensive, such as the now very popular smart speakers. Because we can put data analysis and processing on the cloud, we have a relatively high price/performance ratio."
Song Jiqiang said that many devices are now equipped with AI capabilities, and some devices may not have the ability to sense and perform initial processing. Without the support of 5G networks, the large-scale deployment of AI is very limited.
Commercialization of AI requires 5G network support
"When AI is deployed in a large scale in the application scenario, 5G is also crucial to the landing of AI." Song Jiqiang said, "The applications of AI are usually end-to-end scenarios. Data collection is at the front end, and data processing Value-added in the cloud, the value-added results must be returned to the front-end, in order to improve the front-end equipment processing power and processing results."
Song Jiqiang said: "The ability that AI has now, at least in some areas, has broken through the bottleneck of user experience and can have a good commercial effect, but if the network cannot keep up, user experience will hit a relatively large discount."
He used the smart speakers that some people can now experience to analyze.
The great highlight of smart speakers is the realization of human-machine dialogue. In this experience, unless the machine can put all the speech recognition, semantic analysis and processing, and the final TTS, text-to-speech conversion, are placed in front-end equipment (here refers Smart speakers) can be independent of the Internet. However, smart speakers are often pursuing cheaper AI while pursuing strong AI. However, smart speakers must be placed in the cloud with strong AI capabilities. They can continuously upgrade the vocabulary and vocabulary through the cloud to improve the ability of multilingual language adaptation. . But in the cloud, if the network is not good, this kind of dialogue is often interrupted, and smart speakers are always looking for a network.
"Now smart speakers in China can be regarded as 'hundred boxes of wars', everyone access to voice services is nothing more than several service providers, as long as the network can guarantee that their cloud service is certainly better." Song Jiqiang said. “Now speakers are sold all over the country and we want to buy smart speakers that are easy to use and inexpensive. The network must keep up. Although the smart speakers are simple voice interactions, the amount of data is not large, but the delay of the network is very demanding and must be continuous. The transfer of voice to the Internet reflects the demand for 5G networks."
Complex AI applications that are not yet experienced, such as driverless driving and remote driving, need to transmit all the visual data in the car to the remote driving console. The control information of the control console is also transmitted back to the vehicle. .
Song Jiqiang said that in this kind of application, uploading the video data in the car requires that the bandwidth of the network must be large, the reliability requirements are not severe, and it is not relevant to lose a few packets. The key is to pass most of the information. The scenario requires a large throughput of 5G; the control signal is then transmitted back to the car. At this time, the reliability and real-time performance of information transmission must be ensured. This scenario requires a low latency of 5G and high reliability, so the driverless Remote driving is an AI application that can be done well after 5G deployment.

More:Tenco