Imagine an autonomous car that is moving fast and suddenly encounters an obstacle in front of it and needs to brake urgently. At this time, it is impossible to send the relevant information to the cloud for processing and feedback. Otherwise, a crash may have occurred by the time the braking command is issued. .
Similarly, in industrial scenarios such as steel rolling, the rolling mill needs to rely on multiple servo motors to drive the rolling process in concert, and the motors cannot have the slightest deviation, otherwise the entire steel plate may be scrapped. In order to ensure the rolling accuracy, it is necessary to monitor whether the motor is running normally with the data collection frequency in milliseconds, and it is obviously too late to upload the real-time data to the cloud.
In addition to latency, privacy and security have become a growing concern for individuals and businesses in recent years. For example, in the traditional face recognition scenario, the smart terminal needs to upload the collected facial information to the cloud server for processing, thus laying a hidden danger for data leakage or abuse. For enterprises, commercial privacy is even a matter of life and death, so most enterprises are reluctant to put core data on the servers of cloud platform vendors.
Due to the need for latency and privacy, the concept of edge intelligence has been increasingly recognized in recent years. Seeing such a broad market opportunity, players including Internet manufacturers, IT companies, and industrial giants have entered the field one after another, and Edge AI-related solutions are also blooming in all walks of life.
However, it is easy to do edge intelligence, but difficult to do well, and it is even more difficult to create considerable business value with edge intelligence.
At the previous Lenovo Innovation and Technology Conference 2021, Lenovo Brain Edge AI was officially released. Recently, the Internet of Things Think Tank and other media, Dr. Fan Jianping, vice president of Lenovo Group and head of the artificial intelligence laboratory of Lenovo Research Institute, and Wang Lei, general manager of Lenovo’s commercial IoT business, conducted in-depth communication on the composition mechanism, industry value and layout ideas of Edge AI.
Through their sharing, we know that in order for Edge AI to truly empower thousands of industries, we need to step on several “big pits” in technology, scenarios, and business, and in particular, we must abandon the behavior of “holding a hammer to find a nail”.
Stepping over the pit of “technology”: from cloud to edge
In layman’s terms, traditional cloud computing is very similar to the water supply model of a water plant. Although computing resources can be accessed on demand by turning on the “faucet”, there are also many problems in the “centralized unified water supply model”.
First, the increase in water consumption per household (a surge in data volume) makes the water plant (cloud) somewhat overwhelmed; second, it takes a certain delay (time delay) for water to flow from the water plant to the faucet; finally, once the water supply plant appears The problem will affect the operation of the entire water supply network (privacy security). Edge intelligence, on the other hand, is much like installing an “emergency tank” close to a home faucet to deal with cloud load, latency, and privacy challenges.
According to Dr. Fan, the deployment of Edge AI needs to pre-train the model in the cloud first, and then distribute it to the edge side, which involves the first technical problem – when the pre-trained model in the cloud is When the model is deployed on the edge, the pre-trained model and the edge device may not match perfectly.
To put it into perspective, the cloud “water plant” has a large reserve, while the edge side “water tank” has a small reserve. The process of deploying Edge AI requires operations such as model quantification and compression.
As for the matching of models and hardware, it is reported that, based on the intelligent recommendation algorithm, Lenovo can now provide customers with the optimal AI model and hardware-matching solutions through the Edge AI platform, and make good adaptation of hardware and software. , so that the real synergy between software and hardware can be realized. As a result, customers can realize the adaptation of software and hardware through simulation software without using real hardware in the development process.
When the model is deployed to the edge, the second technical difficulty comes again. This is because when scientists pre-train models in the cloud, all they collect is historical data, but the actual application scenarios are ever-changing. That is to say, the training data and test data may be completely different, and the training environment of the model may be completely different from the actual operating environment – how to update the model on the edge side becomes the key.
Obviously, any model update requires data as support, but unfortunately, it is often difficult to collect enough effective sample data on the edge side. Taking the defect detection scene of a smart factory as an example, we cannot expect 80% of the production line to be waste products, if this is the case, the factory would have closed down early. The actual scrap rate may be only one in one thousand or even one in ten thousand, so the number of defective samples that the model can get is very small. In this case, AI must be able to learn from a small sample like a human.
To solve this problem, Lenovo innovatively combined “data enhancement” and “meta-learning” to create a “small sample lifelong learning technology”.
To give an easy-to-understand example, learn to drive. For example, you can drive a small car, but need to learn to drive a big truck. Although there are many differences between the two, if you can drive a small car, it is easier to learn to drive a big truck than to learn from scratch, because there are some basics you can learn from. This is the principle of meta learning. By learning from the similarities of different tasks, we can generalize the learning ability at the task level and improve the ability of the model to continuously adapt to new tasks.
In addition, Lenovo has expanded the sample capacity, enriched the data distribution, and further improved the learning ability of the model through methods such as original sample space enhancement, data expansion in the model feature space, and style transfer, thereby realizing lifelong learning.
Stepping on the pit of “scene”: abandon “finding nails with a hammer”
Solving technical problems does not mean sitting back and relax, and even many technology companies still fall into the misunderstanding of “holding a hammer to find a nail” – that is, using technology to find a scene, which is obviously a behavior that puts the cart before the horse. . There are many typical application scenarios in the industrial field. As long as it can effectively solve the pain points of customers and help them improve quality and efficiency, there is no need to worry about no one paying.
In terms of scenarios, Lenovo naturally has an advantage. As an intelligent manufacturing enterprise, Lenovo has built a global production base layout, and the group has 35 manufacturing plants around the world. In the past few years, Lenovo’s factories have been committed to promoting the automation and digitization of production lines, and accumulated rich practical experience, which is the best soil for Edge AI to land.
Dr. Fan also shared some practical cases of Lenovo.
When Lenovo Shenzhen factory produced servers for Microsoft, a basic requirement of customers was that no outsiders could enter the production area at will. The traditional method is to rely on manual monitoring at the door, but this method is obviously time-consuming and labor-intensive, and people will always be tired and need to rest. Edge AI’s solution can easily replace manual labor, monitor whether unauthorized personnel enter in real time, and because the data is kept locally, it will not violate the privacy security of employees.
In addition to Lenovo’s own factories, more and more customers are benefiting from it. For example, in order to ensure the smooth operation of Goss China, a leading enterprise in the printing industry, Lenovo only sold PCs before, but later provided edge computing equipment and the Leap IoT platform, etc., and realized the remote cloud management of Goss China, which reduced its equipment failure rate. By 50%, the travel cost of Goss employees has been reduced by 65%, and the customer satisfaction of after-sales service has increased by 80%.
“It was historical opportunity that chose Lenovo, not Lenovo making the choice.” Dr. Fan described Lenovo’s layout of edge intelligence.
According to Gartner’s forecast, by 2025, 75% of data will be generated at the edge outside the data center and cloud. At the same time, the edge computing market will exceed trillions in the future, becoming an emerging market that is on a par with cloud computing. In this context, Lenovo’s entry into edge intelligence is just at the right time – this is the “time of the day”.
Secondly, as a manufacturing company, Lenovo has many application scenarios for itself and many application scenarios for its customers. Lenovo has accumulated a lot of insight and experience in the process of serving customers – this is “the right place”.
Finally, the team of scientists represented by Dr. Fan has always focused on the research of AI technology. As Dr. Fan said, “We are not the smartest group of people, nor the highest-configured group of people, but we do always take customers as our responsibility. .” – this is “people and people”.
“The right time and place” gives Lenovo the confidence to do a good job in Edge AI.
Stepping on the “commercial” pit: cracking the fragmentation problem
Does having the technology and finding the right scene mean all is well? The answer remains no.
Even the same equipment may have different degrees of deterioration under different working conditions, not to mention the thousands of industries that are “interlaced like mountains”. The scene of edge intelligence is highly fragmented, and the requirements are very complex, which changes with customization, resulting in insufficient versatility and reusability of many edge solutions.
“If we do it case by case, it must be a dead end.” Dr. Fan said bluntly, “Lenovo’s solution to the fragmentation problem is that 80% of the underlying technology is implemented on the platform, and the remaining 20% is the scene. things that melt.”
From this point of view, Edge AI is more like an application store, which contains various algorithms and models, and customers can choose flexibly, greatly reducing labor costs such as development and design.
For customers who have a certain understanding of AI technology itself, Lenovo provides basic modules such as personnel detection, traffic flow monitoring, and speech recognition. Developers can quickly combine their own solutions by dragging and dropping like building blocks; for those without AI background Knowledgeable customers, integrators only need to input their own budgets and scenarios into the platform to form constraints, and the platform can still automatically generate corresponding solutions.
In other words, what Lenovo provides is only the basic capabilities at the bottom, and the applications at the top need thousands of customers in the ecosystem to build according to their needs.
Wang Lei explained how to move from customization to generalization from another perspective: “In the beginning, we were working on a project, and some tools would be generated in the process of doing a project. When such tools are used in more projects, the tools will It will become a general product. If the product is further iterated, it will become a platform. Although the customers are different, we will abstract the common things to form a platform and ecology.”
At present, the Lenovo Edge AI platform has been commercialized through Lenovo Enterprise Technology Group, and powerful AI products and solutions such as LiCO AI and AI all-in-one machine have been launched, which are widely used in artificial intelligence data centers, intelligent manufacturing, and industrial Internet of Things. , smart cities, smart retail, smart speakers, and smart homes. In the field of autonomous driving, pedestrian recognition, vehicle recognition, and signal light recognition can be achieved.
Epilogue
Intelligent transformation is not a one-time event, nor can it be turned around by itself. Relying on the commercial IoT ecosystem, the power of joint industry partners, and its own advantages in hardware customization, through the edge AI platform algorithm capabilities and powerful service system, Lenovo can help users achieve faster, better, and lower-cost business models. and technological forms, accelerate the edge intelligent transformation of all walks of life, and then fully leverage the new era of digital intelligence!
The Links: MBM300HR6HY NL6448BC33-63C