Abstract: This article describes how to deliver on the promise of edge AI benefits and work with the open source community on edge AI landing and commercial closed loop initiatives.

This article is shared by Huawei Cloud community “Huawei Cloud: Discussion and Research on the Implementation of Edge AI Solutions”, written by Huawei Edge Cloud Innovation Lab (ECIL).

No fixed source for industry’s actual business and research needs?

Are real business data sets and related algorithms difficult to obtain?

Industrial edge AI systems take time and effort to build?

Hacks can’t find target landing scene?

This is your chance to make your voice heard in academia!

To understand the edge of AI, please click on the portal directly: www.wjx.cn/vj/m9RDmtT….

Edge AI technology trends emerge

Machine learning on the cloud is a traditional and widely known approach, based on cloud-side large-scale computing power, most large cloud platform providers already provide machine learning services. However, the data needed for machine learning is often generated not directly from cloud platforms, but from sensors, phones, gateways and other edge devices.

With the widespread use and performance improvement of edge devices, it has become an inevitable trend to transfer some tasks related to machine learning to edge, that is, edge AI technology. It can even combine cloud computing power and edge data to complete machine learning tasks at the same time. In 2018, VMware released a framework for extending the cloud environment to the edge. Microsoft is also investing $5 billion in the Internet of Things, in addition to Azure cloud, because “the Internet of Things is finally evolving as the new edge of intelligence.” Garner predicts that by 2022, 50 percent of enterprise-generated data will be created and processed outside traditional centralized data centers or the cloud, up from less than 10 percent in 2018, such as inside a factory, on an airplane or oil rig, in a retail store or medical device.

Edge AI technology challenges

Machine learning services need to respond quickly and process locally generated data at the edge in the process of converting data generated at the edge into knowledge. In the process of landing, we found that in the era of edge cloud connecting massive edge nodes, as the distance between AI service and edge users is shortened, part of the original technical challenges of general AI become more acute in edge scenarios. Four challenges are summarized here:

1. Resource constraints: Compared with cheap, on-demand cloud resources, side resources, including computing equipment, power supply equipment, deployment site area, AI development environment, are often limited or heterogeneous. Side service framework processes need to cope with and be compatible with various situations, resulting in higher construction and maintenance costs.

2. Data islands: There is natural geographic distribution at the edges. In industrial applications, AI algorithms often face problems such as data sharing failure, data privacy protection, and even network bottlenecks, which lead to natural geographic segmentation of data sets and the inability of AI algorithms to efficiently and accurately share the data of each edge node. In the traditional centralized AI mode, the performance of various AI systems (including convergence speed, data transmission amount, model accuracy, etc.) decreases in edge scenarios.

3. Small samples: there are usually only a small number of samples in a single edge, especially in the initial stage of side service startup, cold start is common. At the same time, it is difficult to annotate a large number of unstructured samples on the side, and the number of annotated samples is low. This results in the failure of convergence or poor accuracy of traditional statistical machine learning methods driven by big data.

4. Data heterogeneity: There are multiple features, models or annotation distributions in the data set, which directly results in a large difference between the statistical distribution of edge test samples and the training set (also known as non-IID or OOD), resulting in a significant decline in the performance of the general AI model in different situations at different edges. For the same tenant, various services often lead to complex and diverse algorithms and data of different input and output (also known as long tail algorithm or long mantail data). In this case, the process of edge cloud collaborative AI service framework needs to deal with and accommodate corresponding business data at the same time, and achieve efficient resource scheduling through unified tradeoffs.

The following describes the scenario where energy saving parameters are recommended for building air conditioners:

1. Description: The cooler has multiple sets of parameters that can be adjusted. The key to energy saving is to predict the energy efficiency ratio of the cooler under different parameter combinations.

2. Demand for edge intelligence:

A) When the new park system is launched, it needs to be equipped with the capability of side cold start-up to achieve fast delivery;

B) Local customization and automatic closed-loop of park system: data is collected online by edge cloud service and the model is continuously iterated;

C) The intelligent service of devices in the park is offline and autonomous

3. Technical Challenges:

A) Limited resources: The data storage and processing capacity of the devices at the side of the park is limited, and the machine learning service is easy to run out of time while supporting multiple system services, and the local data can only be saved for several months.

B) Data island: different building control and even power system cannot communicate with tenants.

C) Small sample: it takes time to accumulate data when the new park system is enabled. It is not feasible for all parameters combination to run sampling under all working conditions.

D) Heterogeneous data: equipment models in different parks differ greatly, so there is no single universal model. Affected by working conditions, life and so on, the model will change gradually with use.

From the perspective of service application, the current edge AI has the following characteristics:

AR, VR, interactive live broadcasting, video surveillance and other multimedia industry scenarios based on human-computer interaction mainly use unstructured data. Unstructured data refers to data that is difficult to be converted into numerical values or unified formats for semantic parsing by information systems, such as images and texts, and is usually directly processed by human beings. The deep neural network method is mainly used. The most critical part of the four challenges lies in the small number of annotated samples and the disproportionate side resource limitation in complex systems due to the large amount of unannotated data.

Industry, energy, finance and other industries based on traditional electronic information systems are mainly structured data. Structured data refers to data in numerical values or uniform formats that facilitate semantic parsing by information systems, such as database tables, and can be directly processed by information systems. Non-deep neural network machine learning algorithm is mainly used, and its algorithm modeling methods are diverse and highly relevant to business. Among the four challenges, the most critical part lies in small samples on the side, data islands across the side, and service reliability and even interpretability under heterogeneous data.

The KubeEdge community has been paying close attention to edge AI-related challenges. KubeEdge is the industry’s first cloud native edge computing framework and the only incubation-level edge computing open source project within the Cloud Native Computing Foundation. KubeEdge has 800+ contributors and 60+ contributing organizations worldwide, with 4.5K +Stars and 1.3K +Forks on Github. As the only AI Special Interest Group in KubeEdge, SIG AI is dedicated to making AI applications work better on the edge, focusing on edge AI technology discussions, API definitions, reference architectures, open source implementations, and more. In view of the above four challenges, the edge intelligent platform Sedna and its features of collaborative reasoning, federated learning, incremental learning and lifelong learning across the edge cloud have been opened.

Research landing Challenge

At present, various teams in the academic field are implementing their plans and transforming their achievements into the industry. Many teams have encountered a variety of challenges, such as difficult data sets to access, generic solutions that don’t fit a particular business, and a lack of commercial success stories. Technology alone is not enough to complete landing and industrial transformation.

Edge right now, in order to make more friends in the field of AI more complete technology research and development to the ground and economically business closed loop, we’re ready to start the landing challenge research, and combines the latest technology trend of industry focus on community content optimization, the final assembled the open source community strength for the intelligent algorithm of edge developers, service deployment, provide resources and platform to help marketers three roles. The community will bring together 30+ vendors and developers in SIG AI to provide open source data sets, open source preprocessing and feature algorithms, and AI tools that the industry desperately needs.

As a community that believes in open source culture, we always emphasize “Best ideas win”, and the development of the field is no exception. In order to present a more competitive and creative proposal, we hope to know that every friend who cares about open source and edge intelligence has difficulties in the implementation of edge AI related solutions, and then select the Best idea to optimize the community content, so as to present a resource-sharing community ecology reflecting the spirit of open source.

If you have encountered any problems during the landing of edge AI, and if you are willing to contribute to the development of edge AI technology and industry, you are welcome to make fun of it in this questionnaire! Your opinion is vital to the optimization of the open source community. Ridicule is power! Please click on edge AI Landing Challenge to investigate the portal and submit your answer: www.wjx.cn/vj/m9RDmtT…. . The questionnaire was all multiple choice questions and took about 3-5 minutes.

After you complete and forward the questionnaire:

1. After the survey, we will receive the survey report, and each team will have a deeper understanding of the difficulties encountered in the implementation of edge AI solutions and industry transformation. SIG AI of KubeEdge community will continue to help solve problems based on research reports, and have the opportunity to turn crisis into opportunity and technology into productivity.

2. Participating in the questionnaire survey, you will have the opportunity to obtain cutting-edge books in the edge AI industry as a souvenir (30 copies in total, equal contributions are made on a first-come-first-served basis)

Excellent comments: We will choose the best feedback in the answer sheet to issue souvenirs (maximum 10);

Active promotion: forwarding 5 communities and sharing moments with more than 10 likes and recording them to the community assistant (above two-dimensional code), we will issue souvenirs to active promoters (maximum 10 copies);

Enthusiastic participation: Fill in the questionnaire and leave your email address, and we will randomly select questionnaire participants and distribute all the remaining souvenirs;

Edge AI related frontier books

Scan the QR code, send the code “KubeEdge SIG AI” to join the discussion group, get the latest progress of SIG AI, technical dry products

SIG AI KubeEdge community thanks huawei, China Telecom Research Institute, Cient Technology, Harbin Institute of Technology, Nanyang Technological University, Sun Yat-sen University, China University of Technology, Beijing Jiaotong University, Hong Kong Polytechnic University, Wuhan University, Shanghai Jiao Tong University, University of Science and Technology of China and other community members for their efforts in the design process of the questionnaire!

(Research deadline: 23:59 PM, Oct 30, 2021)

Click to follow, the first time to learn about Huawei cloud fresh technology ~