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

Last year, we sensed that interest in AI was close to frenzy, so we launched a survey on AI apps. After the survey was completed, we analyzed the results and concluded that the AI industry was in a stage of rapid change, so we launched another survey to find out which industries AI is mainly used in right now.

The investigation took several weeks and 1388 responses were received. The latest survey tells us the extent to which AI has been used in the enterprise (from prototyping to production), as well as the popularity of specific technologies and tools, challenges faced by users, and other issues.

Let's sum it up:

Key findings:

About 85 per cent of respondents said they were evaluating AI or applying AI to production. Only 15 per cent of institutions have not taken any action on AI.

More than half of the organizations interviewed are "mature users of AI technology", in other words, they are using AI to complete analysis and production tasks.

Supervised learning is the most popular machine learning technology among mature AI users, and machine learning is the most popular technology among those institutions that are still evaluating AI.

In AI applications, the lack of machine learning and AI skills is not the biggest obstacle. 22% of the respondents believe that the lack of institutional support is the biggest problem.

Very few enterprises use formal governance controls to support their AI activities.

AI is still developing. Many enterprises are evaluating AI and experimenting with AI, but in application, it is mainly used for "production deployment" (production deployments). This development is still in its infancy, and companies still have a lot of work to do to make AI stronger. In any case, users still have a lot of work to do to build a stable AI production line.

Interviewee distribution

Respondents were surveyed on behalf of their organizations and companies, who came from 25 industries, accounting for 17% of respondents from the software industry. Judging from the sample interviewed, the proportion of "technology industry" is not very high, and "computer, electronics, and hardware" is also an important part of the technology industry, which accounts for less than 7% of the sample. As for the "other" category, it includes 12 industries, accounting for 22 per cent of respondents.

Data scientists are dominant and corporate executives are representative.

About 1/6 of the respondents were data scientists, with executives (such as executives, vice presidents, CxO) accounting for 26% of the sample. According to the survey, respondents do have a strong bias towards data, with 30 per cent of respondents being data scientists, data engineers, AIOps engineers, or people who manage them. About 3/4 of respondents claimed that their work was related to data. More than 70% of all respondents are engaged in science and technology jobs.

Regional distribution

About 50% of respondents are from North America, most of them from the United States, with 40% from the United States. This was followed by Western European respondents, accounting for 23 per cent, and Asia accounting for 15 per cent. South America, Eastern Europe, Oceania and Africa accounted for 15%.

Analysis: the status quo of AI applications today

More than half of the organizations represented by the respondents have entered the "mature" stage of AI applications, that is, AI analysis, production, and about 1/3 are being evaluated. Looking at the results of last year's AI survey, 54 per cent of respondents said they were evaluating AI and only 27 per cent had entered the "mature" stage. In the 2020 survey, only 15 per cent of respondents did not have any AI programmes.

About 85% of organizations are using AI, most of which are used in production, which is a bright spot of concern. It seems that the experimental AI project developed in 2019 is beginning to bear fruit this year.

What is the main use of AI? About half of the respondents said they used it in R & D, followed by IT, with 1/3 of respondents choosing it (please note that you can choose more here).

Among the options, there are two "functional departments" of the enterprise, namely, "Marketing / Advertising / PR" and "Operations / facilities / fleet Management", each of which is selected by about 20% of the respondents. From this, it can be seen that the interviewees saw the value of AI in the functional departments.

Challenges in adopting AI

In most organizations, the acquisition and retention of AI special skills remains a major obstacle to AI applications. This year, more than 1/6 of respondents said recruiting and retaining AI talent was a major obstacle to AI apps. In 2019, only about 18% of respondents thought so.

But the lack of highly skilled personnel is not the biggest obstacle. In 2020, about 22% of respondents thought that lack of institutional support was the biggest problem, and many people thought so in 2019. In 2019 and 2020, among the major obstacles to AI application, "lack of highly skilled people, it is difficult to recruit the necessary talent" ranked third, "it is difficult to identify appropriate business use cases" ranked second, about 20% of respondents chose.

AI/ML skills shortage

We also asked respondents which machine learning (ML) and AI skills were most lacking in their institutions. Fifty-eight percent of the respondents believed that they lacked ML modelers and data scientists the most, and chose the most. Understanding and maintaining a series of business use cases ranked second, with about half of people choosing. Another 40% of respondents chose "lack of necessary skills in data engineering". Finally, about 1/4 of people chose "institutions lack computing infrastructure skills".

There is one interesting thing about this survey: the results in 2020 are strikingly similar to those in 2019. The problem in 2019 will still be a problem in 2020, and the proportion will be basically the same. In 2019, 57% of respondents chose "lack of machine learning modeling and data science experts", compared with 58% this year. These skills are essential to the organization and are not easy to solve. For example, data scientists are "mixed species". They need not only profound theoretical and technical knowledge, but also practical business knowledge oriented to specific areas.

Unfortunately, many data scientists lack practice, so institutions can only train new data scientists in their work. That is why so many respondents believe that a serious lack of people who understand and maintain specific business use cases is a big problem, rising from 47% in 2019 to 49% this year. Data scientists use their specific expertise to find suitable business use cases for AI. Machine learning modelers accumulate knowledge in practice and then supplement technical knowledge with this specific business knowledge. Both types of practitioners need to improve their soft skills in the process of teamwork, listening and empathy. Ascension takes time and is the result of experience.

Manage AI/ML risk

We also asked the interviewees this question: what risks need to be controlled when you build and deploy the ML model (choose more)? The survey results show that all organizations, especially those in the "mature" stage of the AI project, all agree that they should be alert to the inherent risks of using ML/AI technology in their designs.

In "mature" institutions, about 2/3 of people choose "unexpected outcomes / predictions", which is the largest single risk factor. Among "mature" institutions, the second single risk factor was "controlling the interpretability and transparency of the ML model", with about 55 per cent choosing; as for those "under evaluation", fairness, prejudice and moral hazard ranked second, with 40 per cent choosing. Fairness, prejudice and moral hazard ranked third in the eyes of "mature" institutions, with 48 per cent of respondents choosing. Compared with the companies being evaluated, "mature" institutions are more willing to examine model degradation, so AI model degradation is the fourth largest risk factor in "mature" institutions (about 46 per cent of people choose).

There are nine options given in the survey, and because it is multiple choices, respondents can choose all of them. We found that 41% of respondents chose at least four, and 61% of respondents chose at least three.

Supervised learning becomes dominant, and deep learning continues to heat up.

Among all AI users, supervised learning technology is still the most popular ML technology. More than 80 per cent of mature users already used the technology in 2019, when about 2/3 of respondents were still in the evaluation stage. By 2020, about 73% of mature AI activities have used supervised learning technology.

It should be noted that in 2020, for those institutions in the evaluation stage, deep learning technology has replaced supervised learning technology as the most popular technology. The data speak for themselves: 55 per cent of institutions that are evaluating AI say they are using deep learning technology and 54 per cent say they are using supervised learning technology (multiple choice). For mature AI users, about 66% of respondents said they were using deep learning technology, ranking second, followed by supervised learning technology.

The utilization rate of ML or AI technology is higher among mature recipients than the organization being evaluated. In other words, the two are very different. For example, about 23% of AI activities among mature recipients use transfer learning (Transfer learning) techniques, compared with only 12% of evaluators. In addition, the Human-in-the-Loop AI model is more popular among mature people.

What tools do you choose in your work? 78% of respondents chose at least 2 ML technologies, 59% chose at least 3, and 39% chose at least 4.

Leading tools are becoming more and more important

TensorFlow is still the most popular tool in AI work. In 2019 and 2020, about 55% of respondents chose it, and the proportion remained the same. TensorFlow's strong position also tells us a fact: deep learning and neural networks are becoming more and more popular.

The most popular AI development tool in 2019 will still be popular in 2020. Of the five most popular AI tools, 4 are either based on Python or dominated by Python. TensorFlow, Scikit-learn and Keras remained stable, while PyTorch share increased to 36 per cent.

Data governance is not a top priority.

More than 1/5 of the organizations surveyed claimed that they had introduced formal data governance processes and / or tools to support or complement AI projects.

26% of respondents responded that formal data governance processes and / or tools would be introduced by 2021, and 35% are expected to do so in the next 3 years, which is good news. There is also bad news that AI recipients believe that data governance is a supplement, not an essential element.

Data traceability (Data provenance), data inheritance (data lineage), consistent data definition, rich metadata management and other good data governance techniques will be integrated into the AI project, not superimposed on the AI project.

We can think of data governance as the "observability" part of the software development process: it is difficult to renovate an existing system to make it observable, but it is easier to incorporate observability into the system. Similarly, it is easier to build data governance capabilities for a system or service, but it is more difficult to add afterwards.

Summary of key points

Looking back at the survey report, we summarize some key points that organizations can refer to when promoting AI projects:

If you don't have a plan to evaluate AI, it's time for you to catch up. There are many open source tools, libraries, tutorials and others on the market, as well as approachable general-purpose languages (such as Python), the barriers to entry into AI are greatly reduced. Most enterprises are experimenting with AI, so it is risky to lag behind.

The AI project is in line with the dominant trends in software architecture, infrastructure, and operations.

You can look at it from a broader perspective: AI is used everywhere, not limited to R & D and IT. In the survey, many respondents claimed that they used AI in customer service, marketing, operations, finance and other areas.

To cultivate an organization, it is not just a training model. In the process of AI application, institutional support is still the biggest obstacle. If you think AI can help, then you should take some time to explain key issues, such as how to maintain expectations, why to expect, and what to expect.

-- there are risks in the implementation of AI, and now we have a better understanding of risks. It's easier when we explain to executives and stakeholders what they expect to do when implementing an AI project.

Summary

It is clear that AI is maturing, and although many use cases seem primitive, it still matures day by day. Users are taking active measures to control common risks. Both mature and immature users are experimenting with complex technologies to develop their own AI products and services. Users use a variety of ML and AI tools, but there is basically only one language, that is, the widely popular Python. If you want to scale up your AI practice, there are a few more things an enterprise needs to do: solve data governance and data conditioning problems.

Original link: https://developer.aliyun.com/article/764536?spm=a2c6h.12873581.0.dArticle764536.603a66e35ETgGf&groupCode=aliyunmit

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.