The theory of research on artificial intelligence speaks so much, then how can it come to the ground and form a product or service that can bring value to mankind? Regarding the productization of artificial intelligence applications, Sogou CTO Yang Hongtao used search engines as an example to introduce how to use AI technology to make good products.
The current company is thinking or mistaken in AI applicationI found some data on the Internet. U.S. data said that the number of companies and startups in the global artificial intelligence sector has reached 1,000, and the scale of investment in more than 4 billion U.S. dollars. Domestic Ereli also had data in January, saying that there are 100 companies in the country that have obtained the investment amount. This represents a great concern in this area. But the topic I want to talk about today is that
These companies tend to think about many things they do from the perspective of products. This may be problematic. This will be discussed later. We talked about many fields, such as machine learning, machine vision, robots, recommendation engines, personal assistants, voice-related technologies, and so on. When we think from the perspective of products, we find that it is not the same as thinking from a technical point of view.
In addition, this year's Go Battle has led to an anxiety: whether or not the robot will defeat people. We certainly hope that the machine can overcome people in a limited time, but we hope that from a technical point of view we can find that many problems cannot be solved - for example, we know that deep learning cannot be explained now, and it is difficult for us to reason from a mathematical point of view. It is inexplicable. Because these are not interpretable, it will have big problems in some industrial applications. For example, if there is no explanation when there is a car accident, the people’s trust in it will decline.
On the other hand, it is difficult to truly understand human language ability nowadays with the understanding of linguistic competence. In addition, the production and training of big data requires a great deal of cost and user access to be able to use it in training. Now it is also an area that is not effective.
If you look at this issue from a product perspective, I think from these two perspectives:
The first is that when we talk about whether machines can beat people, we think: What is a victory? If we say that machines run faster than humans, they are faster than people. If such an answer is given, they will find that the machine has already beaten people. Today, when we say that the machine can beat people, it means that people directly target the human brain. In part, can it replace your brain and think about decisions.
The second point is whether the product can make a product form that assists people to make decisions. It is important to say that we can produce enough data.
Search Engine, Artificial IntelligenceThe first one was Siri. When Apple released Siri in 2011, it caused a great deal of discussion and enthusiasm about speech recognition. There are Echo smart home products that sell very well in the United States; and AlphaGo, which is not A product has attracted the attention of ordinary netizens and has contributed greatly to technological advancement; then Tesla’s automatic driving, including a field of consumer concern, has caused accidents and caused a lot of discussion.
But allowing me Lao Wang to sell melons. From the perspective of search engine product practitioners, we think that search engines are actually the biggest scene of artificial intelligence.
Why do you say that? The first reason from the perspective of products, search engines in the past so many years, it played a role in the expansion of the human brain, it enhances your ability to solve problems, many problems in our daily life, work is actually through the search The engine to achieve.
I remember one of Google’s founders earlier said, “Our goal is to become the third human brain.†This describes the unit of search engine products. However, today's search engine is based on the product form of key words and search results. Users should think for themselves and whether these ten sets of results meet their own needs.
Actually, the search engine will have to solve the next problem, or if there is a huge application of artificial intelligence technology in the search engine, that is to say, you can't use such a way of thinking by yourself, but I can directly know what you want. This is the direction of work that everyone wants to solve.
The last is that dialogue robots are still not an alternative to search engines. There were only dozens of chat robots on the line, and now it is one thousand. It is a Kaiping band, and a large number of partners have given it enhanced system capabilities. Siri is an example. On-line service was provided in 2011. Apple developed siri's dialogue function and services provided to users. By this year, it discovered that this function was finally gone and could not be continued. Because a large number of users find it difficult to solve practical problems, when users actually use Siri, they are: Who do I want to call, and who do I want to send SMS to? This year, the developer hopes to provide it with a lot of research and development functions, but in fact it can not replace the answers to the problems caused by the massive number of search engine users.
Search Engine = Computing Power + Data + Application
Regarding the relationship between search engine and AI, we know that the development of artificial intelligence in the past decade, more specifically, the development of deep learning depends on these three elements. Today, due to the progress of the Internet, or the progress of the Internet, enough applications, enough information, and enough users to access the Internet, and finally generate enough data - this data allows us to develop better algorithms, and let us have Very good computing power, with a lot of ways to connect computing power at low cost.
From these three perspectives, who has the best three elements? Still look at the search engine industry leader. Although Google has not disclosed the data of its own server, it is estimated by its energy consumption and it has a number of millions of servers. It can be seen that search companies have such enormous computing power. In the morning, Mr. Zhang mentioned "image recognition for cats." The program can easily connect 16,000 CPUs for this training.
Looking at the data again, search engine companies have large enough web pages to serve as the basis for data. They can also do a lot of how to promote the collection of data. Why did Google launch it and cut it for networking? Why does Google have to pull it for free? optical fiber. Everyone knows that Google’s parent company has two medical subsidiaries in it. One way to do this is to use sensors to put contact lenses in the eyes. To collect the data of a sick person and analyze them, they have this. Conditions to get the data.
The future direction of the work: natural interaction and knowledge to calculate natural interactionHow can you let users interact with products in a natural way? Let the search engine know what you want, not a key word to express - it is a natural language sentence, like a person-to-person dialogue, describing what kind of problems I have. This includes not only language and dialogue, but also vision. You can see your face and appearance and read your emotions.
Knowledge calculation
How can I get enough knowledge to get calculations or reasoning? The current progress we can make in this area is very limited.
In both directions, the current search engine products, or similar dialogue systems, service systems, and solutions are still not good enough. So, how to solve this problem from the product?
Can not solve the problem when the data
When a product experience is not good, what should we do? We have data.
Take a real example of how we do our own products for technology:
Everyone knows that sogou input method has been done for a long time. We realized early on that it is very important for users to interact with machines through voice. At that time we did not have the technology of speech recognition. We also found some companies that engage in voice interaction to cooperate. , but did not find a better way of cooperation. So we forced ourselves to do research and development.
At that time we thought of a way to directly call Google's voice recognition interface provided abroad, although it was slow, but it can be used. Its self-error rate is 43%, meaning that if you enter a paragraph of 100 words and 40 words is wrong. However, because of the entrance of functions in a product, a large number of users started to use it. After using it, we began to iterate data and accumulate enough data for users to record voice. In November 2012, we made our own data in less than half a year. As soon as we got on the line, we got closer to and better than Google. As more and more users are used later, and the use of new technologies in the process, the error rate has plummeted. This year's 4% or so error rate has recently continued to decline.
So when you can't beat it, your data is worth raising. What is worth mentioning, what should we do after we have mastered so much data?
This requires finding some data labeling companies to label the data, what each sentence says, and later because of the use of such a feature in the product, the user does not choose the voice recognition sentence is actually the process of auxiliary labeling. If the user did not select the description is wrong, wrong data we give to the data labeling company to mark.
Bread and Raisins: Artificial Intelligence Makes a Good Product Creates ValueA good product using AI technology creates value. There are two aspects to this value: one is to provide the user with a good idea, and the second is to actually produce data and then iterate itself.
Good product: applause + sale + make money
Good products can get the attention of media and users, can get resources, sell products to get users, users can produce data, and can continue to iterate. Not to mention making money, a good product can make money so that corporate R&D can continue to roll.
Examples of products that are not good enough, we are self-criticizing. Two years ago, Sogou had a cloud assistant product. At that time, it was thought that such a conversational situation may be a new generation of interaction and problem solving in the future. So we made the same product architecture and question-answering technology as Siri Foundation. Architecture. However, after the product was launched, we found that the user's usage rate was not high. It took a while to feel fun for the first few days. Then it was no longer needed. Why? I think there are many reasons. The key reason is that it is not good enough to solve new user needs.
Zheng Yu: In the two examples, I think the second example may be more meaningful, because the search space is limited, and the words are relatively short. The first example of a WeChat message may not be the best one, and it encourages us to drive WeChat. Now let us ask a question: Sogou investment and layout of a lot of artificial intelligence projects, what kind of harvest is currently? What's the biggest result?
Yang Hongtao: My answer is a bit like the answer of just KK (co-founder Huang Jiangji). I don't think there is a product that we feel satisfied with, or that we really let people make decisions or help people make good choices. Such a function has not yet been fully achieved. But our products are really going to solve this problem, and then we have to invest. You just said that I've invested very well and gave me an opportunity to advertise. We invested about RMB 180 million in building an artificial intelligence research institute this year with Tsinghua University to promote technological progress. With technological advancement, we have products. With the advancement of applications, our focus in this area is to make a good product experience. Then to accumulate data, it is still only possible to accumulate data. The next step really is to be able to actively provide assistance to people. I think it has not yet been done.
Low Speed and Middle Speed Generators with Low speed engine or middle speed engine
. Low speed diesel generators, 500rpm, 600rpm
. Middle speed diesel generators, 720rpm, 750rpm, 900rpm, 1000rpm, 1200rpm
. Voltage: 3kV, 3.3kV, 6kV, 6.3kV, 6.6kV, 10kV, 10.5kV, 11kV,13.8kV
. Frequency: 50Hz, 60Hz
Low Speed Generator,Middle Speed Generator,1000Rpm Diesel Generator,500Rpm Diesel Generator
Guangdong Superwatt Power Equipment Co., Ltd , https://www.swtgenset.com