An AI expert talk: reuse algorithms, models, and cases, AI Gallery will take you quickly to get started with application development

Abstract: Huawei Cloud Community invited Yan Bo, the person in charge of AI Gallery, to listen to him talk about the original design, classic cases and future plans of AI Gallery.

This article is shared from the HUAWEI cloud community " AI expert talk: reuse algorithms, models, cases, AI Gallery takes you to quickly get started with application development ", author: HUAWEI cloud community selection.

What interesting and practical AI development cases have you seen?

For example, let the characters in the static photos sing, and the animation characters can also be used; for example, by identifying various wild animals and analyzing the population structure, and then implementing protection; or intelligently detecting the standardization of wearing masks to help prevent epidemics...

For these scene-based AI cases, you can HUAWEI CLOUD AI Gallery . Through continuous training, you can also achieve them. AI Gallery has rich and high-quality AI assets such as algorithms, models, data, notebooks, etc. developers can directly reuse these assets to solve the problems of AI application development.

Throughout the entire development process of AI applications, From data collection, labeling, to the construction of algorithm models, each link will produce many reusable AI assets, and the purpose of AI Gallery is to give full play to the utility of these assets and improve AI Development efficiency.

So, how does it gather these AI assets, and how does it maximize the utility of assets to help developers efficiently carry out AI development?

The Huawei Cloud Community invited Dr. Yan, the head of AI Gallery, to listen to him talk about the original design, classic cases, and future plans of AI Gallery.

There are troikas of artificial intelligence: data, algorithms, and computing power. From these three points, what stage has the current AI application development entered?

Artificial intelligence is a field that human beings are constantly exploring and developing. Because the stages have not been defined in advance, it is difficult to answer which stage we are currently in. But there is a very obvious perception that the development of artificial intelligence has made significant progress compared with 10 years ago, and a large number of applications are also increasing.

is the breakthrough of a class of algorithms represented by deep learning in 2012. Prior to this, everyone was more concerned about algorithms. Everyone was using data dimensionality reduction and some classifier schemes for machine learning-related AI development, and the amount of training data was also very small.

At the opportunity in 2012, with the blessing of computing power, and through algorithm and a large amount of data iteration, we saw an order of magnitude improvement in the accuracy of the AI development model. With this magnitude increase, it can apply AI technology in more industries and fields to increase productivity.

However, at this stage, artificial intelligence is not like human beings, which can obtain logical deduction ability through learning a small amount of data. In essence, artificial intelligence still performs fitting and iteration through a large amount of data, so that it can "remember" the data and then do some reasoning, but it does not have the ability to logically deduction. However, compared with the past 10 years, the final accuracy of AI has been improved, and its application fields have also been further expanded.

Looking forward to the future, we will have further breakthroughs in computing power. Coupled with the blessing of the entire algorithm and data, can finally train an AI with higher accuracy and even logic deduction capabilities like humans.

What are the links in a complete AI application development process, and what are the challenges?

There are generally three processes.

first link of is data preparation 1615aa67b0b9e9, which requires data collection, data cleaning, and conversion. Each link has its own challenges. Taking data as an example, there will be policy and legal restrictions in the collection phase, and data islands are difficult to break. In addition, the data must be effectively annotated, which requires a lot of manpower to complete, and the economic cost is high.

second stage of is to model 1615aa67b0ba40, based on the prepared data, select appropriate algorithms, and develop related models. Consider the application scenarios of the trained model. For example, whether the AI application is placed on the mobile terminal or on the cloud server, the two have different requirements for the delay and accuracy of inference. Therefore, in the modeling process of AI development, it is necessary to comprehensively understand the AI application scenarios, and then choose the appropriate algorithm technology architecture. It is different from the academic field that only pursues accuracy or reasoning speed. We need to consider comprehensively, so the challenge is relatively high.

third stage of 1615aa67b0ba5e is the development of specific AI applications based on models. , which focuses on specific application scenarios and cooperates with the development of some IT systems, software, UI interactions, etc. For example, an algorithm engineer is responsible for modeling and development, but when it comes to the application development stage, it may be assumed by the application engineer, and its role is changed. As an application development engineer, I receive a model that has already been developed, but the time delay and accuracy of this model's reasoning may not reach the ideal state. At this time, it is necessary to further optimize it through compression and distillation technology. If the accuracy is not up to date, you have to consider whether the application can circumvent these problems through some clever design.

Is AI Gallery to solve some of the problems in the above links? What is the original intention of its design?

Many development processes are now platform-based. AI development will produce some digital assets: algorithms, models, data sets, and possibly some processing functions and methods. We hope that there is a place where we can deposit and accumulate these things, so that subsequent developers can reuse some of the previous results . This is also our original intention for designing AI Gallery.

When more and more developers share AI assets in various scenarios, the AI Gallery can contain experiments with various precisions in the entire scene. At this time, other developers can also directly use these assets according to the final development scenario.

For example, the three stages of AI development may involve different roles. What if an application engineer wants to get involved in AI development but lacks the corresponding data and algorithm engineers? The trained model is available in the AI Gallery, and the application engineer can use it right away. From this perspective, it can improve overall development efficiency.

What is the primary consideration for developers to choose an algorithm or model? At this point, how does AI Gallery respond?

If you choose data, it is generally based on its industry and field scenarios, to see if there is suitable data, this is strongly related to the field and industry. Currently we provide a mechanism for data sharing. Many developers have shared data sets of open source standard scenarios for everyone to quickly verify their ideas on ModelArts.

In terms of algorithms, developers first consider whether the final model generated by the algorithm is what they want, the input data format of the algorithm during training, the cost of training, the environment in which the algorithm is run, and so on.

In terms of models, we must first clarify whether the application development is deployed on the cloud, on the edge side, or on the end side. This is very important for the final application scenario. The second is the delay of reasoning. For example, the data of medical scenes will be very large. Its reasoning is asynchronous, but some scenes require real-time reasoning, which may require high inference response time. Finally, accuracy is to evaluate the sensitivity of the application scenario to accuracy.

In summary, at each stage of AI development, there are many indicators and dimensions that need to be considered. What we have to do is to standardize these dimensions and indicators, so that developers who publish AI assets can fill in these indicators, so that users can browse, filter, and retrieve them quickly to find what they want.

What are the classic cases of AI Gallery that can be introduced to developers?

For some classic algorithms in the visual field, such as YOLO and ResNet50, the official has done a lot of adaptations, but these algorithms have not settled in this field and industry. Therefore, based on some internal projects, we have also done some AI practical cases. For example, water meter reading, safety helmet detection and so on. These cases may use the same specific algorithm, but applied in different fields and industry scenarios.

In the future, our partners, college teachers, and developers will share their cases together, so that other developers can quickly reproduce them by reading these cases and accelerate the entire end-to-end development.

Here are some classic cases on AI Gallery:

Safety helmet detection, water meter reading recognition, rebar inventory of construction site scenes, playing Super Mario Bros. using PPO algorithm, and playing against Chinese chess AI.

For example, industrial safety helmet testing and water meter readings are all based on cases precipitated by Huawei's projects in the industry. The models produced in these cases can meet the requirements of industrial grade, and it can be directly deployed and used. The only difference is the data. Currently we only provide a sample of data. If you can collect more and better data, the accuracy of the trained model will be very good.

After AI Gallery releases AI assets, what benefits can developers enjoy?

For ISV partners, AI Gallery is connected to the HUAWEI cloud cloud market, so they can put on the cloud market, sell asset models commercially, and directly obtain commercial benefits.

For developers, the current is more of personal achievements and honors. In the follow-up, we are also actively introducing individual developer plans, allowing ordinary individual developers to participate in the entire project, and truly enter the actual combat link, which can not only obtain practical exercises, but also obtain financial returns.

How does AI Gallery help Pratt & Whitney AI?

One is the accumulation of assets and cases . At present, many developers have contributed mainstream open source data sets on the AI Gallery, and others can use them when they directly verify the algorithm. In terms of algorithms and models, the official has also pre-integrated many commonly used algorithms. Colleges and universities are also posting the algorithms of some classic papers to AI Gallery for sharing.

second is the sharing mechanism . Developers can share algorithms and models to the AI Gallery, and then we are trying to adopt some incentive mechanisms to make them more motivated to share.

Third is for the end-to-end case scenario , we have launched the case library. Although there are not many current cases, follow-up Huawei officials, ISVs, partners, and individual developers will continue to summarize and publish project cases developed and delivered, so that the majority of developers can accelerate the application development process by learning these cases.

What is the future plan of AI Gallery?

first direction of 1615aa67b0bf1e is to accelerate the implementation of AI applications in industries and enterprises. is to improve the efficiency of AI development through asset precipitation. The second is the demand square we are working on, as well as the developer’s certification mechanism. By reducing the links in the development process, more developers and partners can develop and deliver AI projects through the AI Gallery, and ultimately help the industry and enterprises solve problems. Accelerate the implementation of applications.

other direction of 1615aa67b0bf44 is mainly for learning and education scenarios. Now we are doing iterative training and development based on a large amount of computing power and data, but the hardware of many colleges and universities may not be able to keep up, and teaching practice on the cloud is needed. Therefore, for the education industry and the learning scenario for individual developers, we also plan to do some things, including teaching courses, essay interpretation and so on.

In the end, we hope to connect these two lines, not only to provide a one-stop solution for teaching, training and learning, but also to allow developers to put the learned knowledge into practice through real delivery scenarios.

Click to follow and learn about Huawei Cloud's fresh technology for the first time~