Shulou(Shulou.com)06/03 Report--

Friends who often surf the Internet must know that recently, public transport in many cities have gradually begun to accept guide dogs.

According to a widely circulated set of data, as of 2017, the number of guide dogs in China is only 116, which is rarer than giant pandas, while about 8 million blind people need guide dogs.

A series of difficulties and high costs such as breeding, training, internship and employment determine that guide dogs are a "luxury" for blind people. Each guide dog costs 120000 to 150000 yuan, and its life span is only more than ten years. Even if it is lucky enough to have one, how should the visually impaired go out after it retires? this is probably after the hot search is gone. It is still a topic worthy of continuous consideration and continuous improvement in a civilized society.

Recently, a blind man in Turkey, K ü r é at Ceylan, based on Arm's latest processor and NPU, created an AI blind stick that may open a window for more visually impaired people.

So what are the technical requirements behind AI's long-term and safe integration of the visually impaired into public life?

AI Blind stick: it's hard to compete with guide dogs.

Whether AI+ blind stick can help blind people to travel smoothly, we might as well speculate on several important working abilities of guide dogs.

First of all, guide dogs need to accurately identify obstacles.

It includes not only avoiding large pits, cars, pedestrians, railings, etc., but also identifying traffic lights and other key road conditions information, in order to achieve the goal of smooth travel for the blind. Friends who are familiar with AI must know that it is not difficult to detect environmental obstacles based on machine vision + camera + sensor. Therefore, in the AI blind stick, K ü r blind at Ceylan implanted map navigation, obstacle detection algorithm, LED warning light, microphone and so on into the traditional guide stick. Through the ultrasonic detector, the obstacles with high 160cm can be detected smoothly.

At the same time, guide dogs also need to guide blind people to avoid obstacles safely.

Guide dogs who are on a mission will wear a small vest with a lever to guide their owners to walk or stop properly. Moreover, seeing eye dogs make their own judgments based on real-time information, and sometimes even "intellectual disobedience". When they find that the order to move forward is unsafe, they will refuse to obey even if the owner asks to keep going.

The blind staff is different, the initiative is completely in the hands of the blind, even if the voice assistant + AI reasoning chip can carry out independent security warning, this pair of "eyes" is very difficult to restrict the activity of the owner, and naturally there is a certain security risk. In case of personal danger due to equipment and technical reasons, the whole society does not have a corresponding plan and preparation for a series of responsibility division and ethical problems.

The important thing is that guide dogs also need to be integrated into the lives of blind people.

After living with the owner for a period of time, the seeing eye dog will be very familiar with the owner's regular schedule. For example, remember his commuting route, behavior habits, frequent supermarkets and friends, and so on. This personalized memory ability, AI through neural network deep learning, can also be achieved.

However, it should be noted that machine learning training often consumes a lot of computing power, which determines that the AI blind stick algorithm can only be completed by uploading data to the cloud, which will inevitably lead to time delay and security risks of information privacy.

As for seeing eye dogs can establish special emotional contact and trust with their owners, help them expand the circle of social activities, and so on, AI blind staff can not be compared with it before the realization of super artificial intelligence.

Generally speaking, the AI blind stick has been able to complete the functions of navigation and obstacle avoidance at the audio-visual level, but it is still not comparable to guide dogs in judgment, reasoning and emotion. Using it in a limited and relatively safe environment (such as office buildings, etc.) may be the initial scenario in which the AI blind stick can be of value.

Therefore, we also need to think about a new question-why the marginal intelligence, which claims to be able to save AIoT, has not changed our lives as scheduled.

Path-finding Fog Computing: the Landing problem of Edge Intelligence

Since its inception, edge computing has been regarded as an excellent assistant to 5Gbps AI + cloud computing. If cloud computing is the "ultimate brain" of Wisdom, then edge computing is a huge "nerve endings" that bear many "subconscious" reactions.

The AI guide stick, for example, is an excellent scenario for edge computing applications. Guide staff to achieve real-time interaction and judgment, such as seeing the traffic lights turn green, can automatically judge that "passable". It is not necessary to upload the street lamp information to the cloud, and the walking reminder will be issued after layers of judgment of the cloud server. This undoubtedly greatly reduces the travel risk caused by delay and reduces the overload of cloud-end computing.

But the edge computing that allows "Yunnao" to be lazy can also help the industry solve the triple contradictions in the process of AIoT pan-intelligence:

The first is the contradiction between calculation and cost.

In order to meet the real-time and availability requirements of terminal AI inference operation, it is necessary to process a large amount of data locally. Either high-performance AI chips are deployed in the terminal itself, which is obviously not realistic in terms of cost control, or enough edge AI is deployed in physical scenarios.

Of course, in order to meet the massive computing needs of AIoT, it is necessary to transform the network pipeline, such as the establishment of 5G edge data centers, the training of high-performance algorithms, and the competition for computing resources such as NPU and GPU, which can not be solved overnight.

The second is the contradiction between instant and power consumption.

For devices such as guide sticks, not only to ensure real-time, but also need to deal with complex AI tasks such as object detection, speech recognition, gesture monitoring, and even face recognition, coupled with the wide range of sensing processing, directly lead to high power consumption. The battery life is only five hours, in other words, it may be difficult for blind people to go back when they go out in the morning and no electricity at night.

Edge computing can filter and analyze the huge data traffic at the terminal, reduce the transmission path from the device to the cloud, and naturally improve the problem of power consumption.

The third is the contradiction between convenience and safety.

Everyone knows that the Internet of things collaboration can greatly improve the portability index of life, but in this era when intelligent door locks and cameras are frequently selected by hackers, data can easily be used by people with ulterior motives. Many enterprises even require that AI must be deployed on their own private cloud, which limits the application of many cutting-edge technologies and increases the difficulty of operation and maintenance.

The solution of edge computing is to put the data processing and storage locally, which can not only protect privacy and security, but also achieve efficient real-time interaction and iteration. In particular, the large-scale application of IoT products such as guide stick, pacemaker and smartwatch, which carry users' life and health information, does not leave the widespread popularity of edge computing.

From this point of view, the AI guide stick is just an example of AIoT innovation. According to IDC's forecast, the number of Internet of things connections will grow to 27 billion and the number of Internet of things devices will reach 100 billion in 2025. It is conceivable that as the edge computing system between the cloud and the end continues to mature in the future, more and more innovations and creations will be excavated to help people with disabilities lead a normal life, help cities prevent micro-growth, and inject AI into thousands of industries.

The future of marginal intelligence still needs to wait for the weather.

The full bloom of marginal intelligence will naturally give birth to huge industrial rich mines and new commercial opportunities. Everyone must have rubbed their hands and wanted to fight hard.

However, it should be noted that although marginal intelligence is the trend of the times, it also has the rhythm and timing of growth, and blind entry may gain nothing.

At present, it seems that marginal intelligence still needs to wait for the full maturity of the industrial environment:

First, the improvement of infrastructure.

As a promising future trend for cloud manufacturers, the software and hardware of edge computing are basically in place. For example, ARM released DynamIQ technology and related processors for artificial intelligence applications, which aims to build distributed intelligence from network nodes to the cloud; NVIDIA's development board Jetson TX2 can also better run deep learning functions on terminal devices.

But this is not enough, edge computing and 5G intelligent network, I am afraid is really glue-like "original match".

On the one hand, the current 4G network construction is generally based on a centralized core network, which is usually difficult to achieve local diversion (Local-Breakout), which leads to a very long physical distance for the data to reach the application side. In other words, low latency requirements cannot be guaranteed by building edge intelligence on top of 4G networks.

In addition, edge computing is not just a simple assignment of computing tasks, making reasonable use of local free time and assigning tasks to different quota computing nodes, all of which require an intelligent network to arrange troops and achieve load balancing. in order to ensure the efficient utilization of each edge node. At this point, 5G intelligent network is also more reliable.

The pace of 5G construction affected by the epidemic, supply chain, etc., will be slower than expected, which further delays the iterative upgrade of edge computing nodes (such as probes, processing equipment, data centers, etc.).

The second is the linkage of industrial application.

Since it is AIoT, it naturally needs multiple edge nodes to cooperate and exert the maximum value of AI through the integration of technology.

For example, when blind people travel with AI guide sticks, telephone poles, speeding vehicles, garbage bins, traffic lights and other nodes share real-time data with edge nodes. Will the AI guide stick be more efficient and feasible than visual recognition solutions to make accurate obstacle avoidance judgments based on these data?

Other nodes can also optimize and influence the city's traffic management as a whole through data sharing, training and mastering travel big data.

At present, this kind of application linkage of edge collaboration is still in the ideal. A more realistic solution is to accumulate and train relevant models through flake updates of smart parks, smart buildings, smart cities, etc., and finally integrate industrial edge intelligence with consumer-grade Internet of things to form ubiquitous intelligence of everything, so that AI can be summoned anytime, anywhere.

The third is the cultivation of development ecology.

Without rules, there is no square, and the future of a comprehensive intelligent Internet of things naturally requires unified standards and norms. But although many cloud vendors have delivered a lot of edge computing tools. But up to now, we have not seen the creativity and creativity of developers explode in the field of AIoT.

For example, in July 2018, Google launched two products for large-scale development and deployment of smart connected devices: Edge TPU and Cloud IoT Edge; Amazon also decided to extend AWS to intermittently connected edge devices at the 2016 re:Invent developer conference. Microsoft's Azure IoT Edge, which also allows cloud workloads to be containerized, can run locally on smart devices from Raspberry Pi to industrial gateways.

Apart from the update iteration of traditional hardware manufacturers, there are few innovations such as the AI guide stick that subvert the traditional functions.

The core reason is that the development threshold is still too high. In addition to the technical threshold for the use of software and the cost of training machine models, there is a lack of integrated software and hardware systems and unified and reliable industry standards, which require developers to pay attention to cross-platform compatibility, heterogeneous data processing, integration of different technologies and ecology, and so on, which undoubtedly consume too much energy and time, deter many developers, and limit the emergence of more creativity.

Cultivating developer ecology from now on may be the key for cloud manufacturers to lead the industry standard, end the chaos and open the competitive position in the future.

The iron law of the science and technology industry is that technology serves applications, while new applications create new market leaders. 4G is to the mobile Internet and AI is to the digital industry.

If the intelligent society is still a ravines and ravines, the primitive jungle is mysterious, but it also has countless treasures to be excavated. Then marginal computing may be the "guide stick" leading to the future.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.