The first article: first acquaintance
Quyc April 27, 2022
Overview of artificial intelligence platform knowledge
First, understand the concept of (what)
Before talking about the AI platform, we can give a general explanation of AI. AI is artificial intelligence, which is a technical science to study, simulate, extend and expand the complex behavior of human brain, including machine learning and computer vision. Then the AI platform is an auxiliary tool to support this complex scientific research, which can complete this series of research intelligently, systematically and automatically.
1)AI platform is mainly for model developers, and provides tools around the life cycle of AI model/algorithm (data collection, data labeling, model construction, model training, model optimization and model deployment).
2) The AI platform is user-oriented, deploying applications around integrated AI services, and mainly performing related operations such as application management.
Second, the product function (How)
Next, the functions of the two classifications of the AI platform will be described one by one.
2. 1 artificial intelligence development platform
1) data annotation platform
The preliminary work of model training includes data import, data preprocessing, data labeling and data enhancement. This part of data work has a strong correlation with big data, and some labeling platforms are even part of big data systems.
For AI tagging, data processing is more intelligent/automated, so some manufacturers have introduced intelligent algorithms (data enhancement) such as data sampling, data splitting, data missing value processing (data preprocessing), automatic tagging (data tagging), image type data defogging, atomization, contrast enhancement, and so on, and it is these functions that support the data tagging platform.
2) Model training platform
Configuring computing power and environment for the content of model training is a common product in AI platform. Because of the high consumption of hardware resources for model training, cloud computing resources are usually rented to complete model training, so many model training platforms are bundled with cloud platforms to complete tasks including load balancing and parallel training.
4) Model deployment platform
Provide tools to deploy models from training environment to reasoning environment (cloud, edge, etc.). ). This function is relatively simple, rarely used as a product alone, but generally used as a functional module of the development platform.
An exception is the deployment platform of edge/ embedded environment (such as Baidu EasyEdge). Due to the complexity of hardware adaptation, Baidu is currently regarded as a relatively independent product.
5) Model reasoning platform
Provide a variety of model interfaces for users to call directly, and generally provide functions such as model call management and interface management. This reasoning platform mainly takes the model as its core competitiveness. Another reasoning platform is competitive in computing power, similar to the cloud platform. Users can gain the ability of flexible scaling by deploying models on the platform.
2.2 ? Artificial intelligence application platform
Compared with AI development platform, AI support platform is more similar to business platform, such as content audit, intelligent dialogue and so on. Around a core algorithm, the universality of this algorithm/capability is improved through configuration.
Explain in detail through the following examples:
Horizontal is the business process of publishing pictures, and vertical is the function of the audit platform. The core issue of the audit platform is the classification of pictures, and the pictures that meet the requirements of the audit policy are restricted.
Third, the core advantages (why)
The advantages brought by AI platform can be considered from the perspectives of users and platforms:
Users: acquire AI capabilities at low cost, improve work efficiency, and meet the needs of rapid business expansion;
AI platform: standardized working tools/processes, which can be solved without customization, improve model production efficiency, reduce working costs, and thus form profits;
But at present, the demand of the platform is greater than the demand of users, which is related to the development history of AI. At present, AI technology is still in the primary stage, and it is more a demand for model training and a process of standardization of industry solutions; User AI still holds a wait-and-see attitude (uncertain to improve ROI, etc. Therefore, it is concluded that it is necessary to continuously strengthen the cultivation of AI capabilities to better meet the needs of market users, and at the same time, it is necessary to continuously guide users to discover the value of AI capabilities, thus enhancing the value of AI platform.
Iv. market situation (where)
The life cycle coverage of some AI development platform products shows that most products actually provide full life cycle functions and provide one-stop solutions.
Competitive product analysis:
Baidu's functional architecture is the most comfortable and logical. Baidu's AI development platforms include BML and EasyDL. BML is a full-process development platform, covering the whole life cycle of AI model. EasyDL positioning is zero threshold development, so it only supports development to data training level. BML's relatively independent data-related functions and edge deployment-related functions are also split into components/small platforms, which users can call independently, improving flexibility.
In Tencent's TI series platform, TI-ONE is positioned as a "one-stop machine learning service platform", but the function of data annotation has not been seen yet, and data processing only provides relatively simple functions of data access and data preprocessing. There are relatively few preset models, most of which are machine learning models and few deep learning models.
The other two platforms of TI series, TI-Matrix and Ti-EMS, are respectively "AI application service platform" and "no service reasoning platform". Individuals prefer cloud services, mainly in service scheduling, capacity expansion and contraction.
Huawei ModelArts also provides development tools for the whole process from data annotation to model reasoning. Among them, the "automatic learning" function module basically benchmarks Baidu EasyDL, providing a model generation retraining level, but the product has not been split according to the demand level for the time being.
Summary: At present, the AI platform has its own emphasis according to the different needs of users, but the training ability of one-stop platform is basically deployed, mainly in three aspects: data, model and deployment;
1) data differentiation: further consistent with the big data platform, providing data collection, cleaning, labeling (automatic and manual) and other functions to solve the pain points of user data.
2) Model differentiation: provide a more powerful preset algorithm for model training, model training for different business scenarios, and optimization for different business scenarios. Secondly, rich instance resources are needed for good connection and collaborative processing with the cloud platform.
3) Deployment differentiation: convenience, quickness, rapid construction and flexible application have become a major difficulty to be overcome in deployment, which is also a very important competitive advantage, saving time and labor costs and facilitating operation and maintenance;