Wunda: Big data will eventually help the machine to have its own wisdom

Forbes website to join the Baidu Artificial Intelligence laboratory Wunda an interview. The article pointed out that the recruit Nawunda, reflects the Baidu hope that through the development of world-class technology, will itself to create the forefront of the world's innovative company vision. In the following interview, Wunda revealed how he will help Baidu achieve this vision.

The following are the main contents of the article:

In May this year, Baidu created the artificial Intelligence laboratory in Silicon Valley and included Wunda as chief scientist of Baidu. As a computer science professor at Stanford University, Wunda was the head of the Googlebrain project and founded the online education start-up Coursera. Let Baidu become internationalized Enterprise's development plan, Wunda lose as the core personage.

Q: How do you have an interest in AI?

Q: There was a time when you lost interest in AI because of the bad progress, didn't you?

A: The broad prospect of AI is that machines will one day evolve to be like humans, capable of accomplishing tasks with their own wisdom. When I first entered Stanford, I didn't think the prospect was feasible, so I was a bit confused. The wisdom of man may depend on some learning algorithm. So, I think maybe we can imitate the human brain, build more intelligence like the brain, and make rapid progress. These ideas have been around for a long time, but Numenta, the co-founder of AI experts and brain-inspired software vendors, Jeff Hawkins, Hawkins, has contributed to popularizing these ideas.

Q: What is your current progress in really realizing these ideas?

A: We are far from true success. We are faced with many problems. One of them is that it is not up to size. At present, our image processing scale is far from enough. Secondly, I'm pretty sure we haven't figured out the right algorithm yet.

Q: But why have people renewed their interest in and attention to AI in recent years?

A: About four years ago, at the end of 2010, we had figured out a lot of algorithms and realized that the biggest bottleneck in pushing artificial intelligence technology to a higher level was scale. If we use the current computer to run the software written in the 1980s, the effect is much better than that of the computer in that era.

So at the end of 2010, I was looking for ways to expand the size of the algorithm in Silicon Valley. Google has a lot of computers, so I started a project in Google that uses these algorithms to build neural networks that are larger than ever. Now, in retrospect, the key to the success of the project is that the instructions the team has accepted are simple: build as large a neural network as possible.

Q: You mean the 2012-year Google Brain Project, when the project's neural network successfully identified the cat's image.

Answer: Yes. Google's neural network has been able to find its own definition of the cat, very impressive. No one has ever told it what a cat is. That's a milestone in machine learning. This is instructive to many companies, such as Facebook, Baidu and so on.

Q: What is the striking point of this project for these companies?

A: Most of the cost-effective applications, so far, are only learning from tagged data. Take speech recognition as an example. Baidu and Google have improved the recognition effect of speech, based on the theory that the depth learning algorithm can accept a large number of transcripts of text-to-speech data. This is the tagged data. So we can train the neural network to predict.

A: In the past I just thought it would be cool to make machines with artificial intelligence. During my high school summer vacation, I was an artificial intelligence intern at the National University of Singapore, tasked with writing neural networks. It is the embryonic form of the depth learning algorithm. I think it's interesting to write software that can learn from self and make predictions.

If we can make computers smarter and better understand the world and the environment, we can make life better for many people. Just as the Industrial revolution liberated us from physical labor, I think there is a huge potential for artificial intelligence, which will allow us to get rid of a lot of repetitive mental work in the future.

In the long run, there is a different kind of depth study, I am very excited about it. It is called unsupervised learning (unsupervised learning), which means learning from unlabeled data, which is closer to the way the human brain learns. The Google Brain Project's identification of cats is an example of this. We've had neural networks watch YouTube videos for a week without any hints. A week later, we did a test to see what it learned. As a result, it learns to recognize human faces, cat faces, and other things. In a way, this is a major advance in artificial intelligence. At present, supervised learning is an important contributor to the economic effect of deep learning technology. Unsupervised learning is not the same concept as it is.

Q: Why do you watch unsupervised learning?

A: One reason is that unsupervised learning is the closest way to learning animals and babies. If you want the current neural network to learn to recognize cars, our approach is to look for 50,000 of pictures of cars and label them as cars, and then enter these tagged data into supervised learning algorithms. And how do children learn to know cars? No parents will get a picture of 50,000 cars. Most neuroscientists believe that the learning of most animals and children is accomplished only by being integrated into the world and experiencing the world. If we can make progress in this area, we can let the neural network system to better understand the image.

The second reason is that we can provide limited data for certain applications that rely on supervised learning, thus limiting its potential for development. For example, in medical imaging, the number of X-ray scans across the country is limited, so that the image data available is restricted.

Q: Now working in the Baidu Laboratory, your focus is not unsupervised learning?

A: It is one of the important issues at the beginning. Unsupervised learning faces more difficulties, predecessors have less successful experience, and do not know what the correct algorithm is.

Q: There is a trend that people tend to study the computing and communication of mobile devices. In addition, sensory data is also showing an eruption. Are these two factors triggering a boom in artificial intelligence?

Answer: There are other factors. The development of large data stems from two trends. First, the increasing depth of social digitization has spawned electronic data that computers can handle. Second, the cost of storage and computing is falling, ultimately reducing the cost of storing and processing all of this data to the extent that it is affordable. If social digitization continues to develop, storage and computing costs continue to slide, big data will eventually develop into a trend.

Q: For the Baidu Artificial Intelligence Laboratory, what are your specific plans in the short or medium term?

A: Baidu has three laboratories, two Beijing laboratories have begun to take shape, and the artificial intelligence laboratories in Silicon Valley are mostly empty, just beginning. We recruit new members very quickly and recruit one member a week, so far we have recruited 6 people. So far they have come from Silicon Valley and have accepted our offer. We are also in contact with a few people outside Silicon Valley, which will take more time. We have a lot of work to do.

Q: Have you drawn a blueprint for the Silicon Valley Artificial Intelligence lab, for example, will you follow Bell Labs, Xerox Parker Research Center, Google or Microsoft?

A: I have spoken to the directors of these laboratories before. I met with Burmack of Sri Labs this morning and talked to a number of people in the field, such as the former director of the Xerox Parker Research Center. I worked at Google X Labs. Long ago, during the early summer of undergraduate and Ph. D., I practiced in T Bell Labs. It's important to keep learning from others and being humble.

Q: What is the most important factor in a successful laboratory?

A: Team culture.

Q: It's important to position correctly from the start. It is difficult to correct if you start off in a biased direction. What do you think?

A: After the initial period, it is difficult to make some changes. The key is what the task is. All the institutions I have worked for are mainly motivated by tasks. Coursera's mission is to educate everyone, so we decided that the best way to achieve this is to create Coursera. Now, the task I am facing is to change the world through artificial intelligence technology. And I have a strong feeling, to achieve this, the most effective way is to join Baidu.

Q: Why choose Baidu?

A: Baidu already has very advanced depth learning technology. Its deep learning laboratory director Yukei is an expert in depth learning. Deep learning of Baidu's core products, such as web search, advertising, speech recognition, optical character recognition and so on, is of great significance. Li has a great passion for AI.

I joined Baidu for three reasons.

Firstly, artificial intelligence is a capital-intensive technology. To make progress, you need support for data and computer resources. Data is more difficult to obtain than computer resources, but both are indispensable.

The second is flexibility. As a large enterprise, Baidu has incredible flexibility. For example, Yukei wants to create a graphics processor cluster, which is quickly put into practice after making a decision.

The third is the enthusiasm of the staff. Baidu's engineers work very hard.

Q: How do you plan to start your future work in Baidu? In addition to Internet applications, will it be involved in the study of robots, unmanned vehicles and other similar products?

A: Initially, we're just going to focus on technology. Usually I tend to take the product as a foothold and then consider the technology. In Silicon Valley, most failures are not due to difficult technical problems, but to the eventual discovery that there is no such thing as an effort to solve the problem.

Deep learning this technology is very different, it has been applied in many products of Baidu. Ai is of great significance to many things. So I don't have to worry about whether our research content is practical for current or future products.

Q: How will you use the technology infrastructure of Baidu? Do you need to build some new facilities?

A: I'm trying to figure out how to use the current infrastructure and development tools to enable a deep learning team to efficiently generate new ideas and test and learn. For example, the implementation of many speech recognition experiments takes about a week. It's hard to learn efficiently if you get feedback on results in a week. If you halve your time, you can double your team's efficiency.

Q: What new technologies might be developed in the future?

A: I hope to be able to make Baidu's current in-depth learning applications continue to improve, such as search, advertising, language translation, optical character recognition and speech recognition.

The progress of technology is divided into two kinds. One is gradual, and this progress will be welcomed. For example, if we can improve the performance of Web search by 5%, many users will benefit from it.

Q: What about the other one?

A: Another technological advance is groundbreaking, and it will advance the emergence of unprecedented new applications. For example, if speech recognition technology is progressing to the point of accessibility to human language, it will create a new mode of mobile phone interaction. Imagine if we were driving, even if the car was noisy, we could still send messages to our friends by typing voice into their phones. This cannot be achieved at this time.

If we can really solve the speech recognition problem, I want to revolve around the voice domain to face the handset to redesign. Perhaps the future of mobile devices in the email application only need two buttons: reply and delete. This is just my idea, not necessarily feasible. But it shows that breakthroughs in some of the core technologies will make a huge difference in people's lives.