Editor: David La Yan

Zhiyuan Xin guide

Andrew Ng, a pioneer of artificial intelligence, was interviewed and talked about his outlook on the future trend of artificial intelligence in 10. He believes that the future technology will shift from hardware to data, forming AI.

Do you feel that you have had enough of your present job and want to change your direction? If you have, you are definitely not alone. However, there are some less radical methods besides joining a big dictionary, such as Andrew Ng's method.

Andrew Ng is one of the most outstanding figures in the field of artificial intelligence.

He is the founder of Landing AI and Deep Learning. Ai, co-chairman and co-founder of Coursera, is an adjunct professor at Stanford University. Previously, he was the chief scientist of Baidu and one of the founders of Google Brain Project.

But according to himself, his focus has shifted from the digital world to the real world, just as the saying goes.

In 20 17, Andrew Ng founded LandingAI, a startup company dedicated to promoting the application of artificial intelligence in manufacturing.

We interviewed Andrew Ng and discussed his data-centric artificial intelligence method, and its relationship with his work in Landingai and the background of artificial intelligence today.

From digitalization to landing

Andrew Ng said that his motivation is industry-oriented. He believes that manufacturing is

Many countries, including the United States, are saddened by the decline of manufacturing. Andrew Ng hopes.

This is a growing trend. According to a survey of 202 1, 65% of the leaders in the manufacturing industry are trying the pilot of AI. It is estimated that the compound annual growth rate will reach 57.2% in the next five years.

Although artificial intelligence is increasingly used in manufacturing, this process is much more difficult than Andrew Ng imagined. He admits that when Landinggai started, it mainly focused on consulting work.

However, after participating in many customer projects, Andrew Ng and Landinggai developed a new toolkit and game manual, which enabled AI to play a role in manufacturing and industrial automation.

LandingLens is committed to helping customers in manufacturing and industrial automation to quickly and easily establish and deploy vision inspection systems. Wu Xiaobo had to adjust his work in consumer software and focus on artificial intelligence in manufacturing.

For example, computer vision driven by artificial intelligence can help manufacturers complete tasks such as identifying defects in production lines. But it's not easy, he explained.

Andrew Ng said that the challenge faced by many companies in the AI field is how to help 10000 manufacturers build 10000 customer systems.

The data-centric approach holds that AI has reached the point where data is more important than models. If AI is regarded as a system with moving parts, it should keep the model relatively fixed and focus on high-quality data to fine-tune the model instead of continuing to promote marginal improvement of the model.

Not many people have this idea. ChrisRé, who leads the Hazy research team at Stanford University, is another advocate of data-centric approach. Of course, as mentioned earlier, the importance of data is not new. There are mature mathematics, algorithms and system technologies to process data, which have been developed for decades.

However, how to establish and re-examine these technologies on the basis of modern artificial intelligence models and methods is a new requirement.

Just a few years ago, we didn't have a long-lived AI system or a powerful depth model of this scale. Andrew Ng pointed out that since he started talking about data-centric AI in March, 20021year, the reaction he got reminded him of the scene when he and others started talking about deep learning about 15 years ago.

Andrew Ng said.

Artificial Intelligence and Basic Models

If artificial intelligence with data as the core is the right direction, how to operate all this in the real world? Andrew Ng pointed out that it is unrealistic to expect institutions to train their own customized artificial intelligence models.

The only way out of this dilemma is to design a tool that enables customers to design their own models, collect data and express their knowledge in their respective fields.

Andrew Ng and LandingAI will realize this through LandingLens, enabling experts in various fields to transfer knowledge through data tags. Andrew Ng pointed out that in the field of production, there is generally not a lot of data for reference. For example, if the goal is to identify defective products, then a fairly good production line will not have so many pictures of waste products to refer to.

In the field of production, sometimes there are only 50 pictures in the world for reference. This is not enough for the existing AI. This is why the focus should now be shifted to letting experts record their knowledge by collecting data.

Andrew Ng said that Landinggai's platform is doing just that. The platform can help users find the most useful cases, establish the most consistent labels, and improve the quality of pictures and labels input into the algorithm.

The key here is. Andrew Ng and others before him found that professional knowledge cannot be defined by an expert. What is defective for one expert may be valued by another expert. This phenomenon is not unique, but it only appears when you have to generate datasets with the same annotations.

Andrew Ng said:

This method is not only meaningful, but also has some similarities. The process described by Andrew Ng obviously deviates from the methods commonly used by AI today, but more points to methods based on management, metadata and semantic coordination.

In fact, people like DavidTalbot, the former head of machine translation at Google, have been conveying the idea that besides learning from data, it is also meaningful to apply knowledge in various fields to machine translation. In the case of applying machine translation and natural language processing, the knowledge in this field refers to linguistics.

We have now reached a new stage, and we have a so-called NLP basic model: for example, a huge model like GPT3. After a lot of data training, people can use these models to fine-tune specific applications or fields. However, this basic NLP model does not really use the knowledge of various fields.

Can the basic model of computer vision do this? If so, how and when can it be realized? What will the realization bring? Andrew Ng believes that the basic model is not only a scale problem, but also a traditional problem. He thinks this can be achieved because many research groups are trying to establish the basic model of computer vision.

Andrew Ng said:

The next 10 year of artificial intelligence

As an insider of computer vision, Andrew Ng is well aware of the steady progress that artificial intelligence is making. He believes that at some point in the future, the media and the public will announce that the computer vision model belongs to the basic model. However, whether it can accurately predict when it will be realized is another matter.

For applications with a large amount of data, such as NLP, the amount of domain knowledge input into the system will decrease over time. Andrew Ng explained that in the early days of deep learning, people usually train a small deep learning model and then combine it with more traditional knowledge base methods in various fields, because the effect of deep learning is not good.

However, as the scale of the model becomes larger and larger, there are more and more data, and less and less knowledge is injected into various fields. According to Andrew Ng, people tend to think that a large amount of data is a learning algorithm. This is why machine translation finally proves that the end-to-end purity of learning methods can be well expressed. But this only applies to problems that need to learn a lot of data.

Domain knowledge does become important when you have a relatively small data set. Andrew Ng believes that artificial intelligence system provides two sources of knowledge-data and human experience. When we have a large amount of data, artificial intelligence will rely more on data than human knowledge.

However, we can only rely on human knowledge in areas where data are scarce, such as manufacturing. The technical solution is to build tools for experts to express their knowledge.

This seems to point to methods such as robust artificial intelligence, hybrid artificial intelligence or neural symbol artificial intelligence, and technologies such as knowledge map used to express domain knowledge. However, although Andrew Ng knew about these technologies and found them interesting, Landinggai didn't cooperate with them.

Andrew Ng also found that the so-called multi-modal artificial intelligence or the combination of different input forms is promising. In the past ten years, people have focused on establishing and improving single-mode algorithms. Now that the artificial wireless city has become bigger and has made progress, it is meaningful to pursue this direction.

Although Andrew Ng was one of the first people to use GPU for machine learning, he doesn't pay much attention to hardware now. Although it is a good thing to have a thriving artificial intelligence chip ecosystem, including established companies such as NVIDIA, AMD and Intel, as well as upstarts with novel architectures, this is not the end.

In the past ten years, most of the focus of artificial intelligence has focused on big data-that is, let's use huge data sets to train larger neural networks. This was promoted by Wu Enda himself.

However, despite the progress made in big models and big data, Andrew Ng said that he believes that the development focus of artificial intelligence should shift to small data and data-centric artificial intelligence.

Andrew Ng said:

References:

https: