/ Home / Events

Career devoted to developing 'electric brain'

 |  | 

 

Zhang Quan left a career as a telecoms engineer to work in a then-unknown branch of artificial intelligence called hierarchical network of concept (HNC). The goal of his team at the Institute of Acoustics is to teach a computer how to "read" books and absorb the knowledge.

Recently Zhang Quan was interviewed by China South Morning Post. Following is the dialogue between Zhang Quan (Zhang) and Journalist (J).

J: How does the HNC determine whether a webpage contains undesirable content?

Zhang: Most of the mainstream content filtering programmes today rely on keywords. They block a page when certain words are found. It is a mechanical process. Quite often innocent pages become victims. Websites that attack "porn video", for instance, are blocked because they use a keyword. HNC goes beyond keywords and makes a computer capable of understanding the context in which the keyword appears. It regards "Porn videos are our enemy" as safe because it knows that an enemy is bad. It can do so because it understands our natural language.

J:  How do you "teach" a computer Chinese?

Zhang: We teach computers like we would a child. We don't treat them as a tool or a machine that can only do what it is told. When we teach a computer human language, it is possible to give them feelings and emotions and the capability to independently analyse what they are told and generate creative feedback. This is not science fiction. Treating computers like humans is the guiding philosophy of our team.

When we teach it a language, we give the computer a chance to simulate how the brain works. Individual words are what children first learn to understand but unlike a child who needs years to build up a decent vocabulary, a computer can absorb millions of words in seconds. So the first thing we did was to compile a dictionary. The dictionary is similar to the ones that rest on our desktops but the difference is that our dictionary translates the natural language words, with each potentially having multiple meanings, into a string of abstract symbols that the computer can understand. Then we teach the computer how to understand a sentence. It's an extremely challenging process. Here is an example: in wo chi fan (I eat rice) chi means "to eat". Wo chi cha means "I drink tea" however. A human brain can tell the different meaning of chi in an instant because it knows the difference between tea and rice and is capable of building a link between them.

HNC enables a computer to "think" in the same way. We call it understanding when a computer can correctly determine the relationship between all the words in a sentence. In our experiments, the accuracy of HNC has reached up to 80 per cent.

The third stage, which is what we are working on right now, is teaching a computer how to find a theme in a paragraph. In future we hope the computer can understand an entire article without human assistance. But it might take several more generations of research.

J:  Why did you choose artificial intelligence as a life-time career?

Zhang: My undergraduate and master's degrees are both in signal processing, a field light years away from what I am doing now. Most of my time then was spent on noise reduction for electric communications such as long-distance phone calls to make signals sharper and clearer. I didn't even have a chance to look into some of the most basic mechanisms of how a computer works.

When I was about to graduate, I looked back upon what I had been doing in the previous few years and found that the field was really lacking the excitement of uncharted areas and there were simply too many people equipped with the same knowledge and skills. So I poked around and found that in the Acoustic Institute a very well-known professor called Huang Zengyang had just established a new discipline called HNC and needed a PhD student. With a wild imagination inspired by science fiction and movies about artificial intelligence, I entered the programme.

Almost immediately I found that most of the things I had supposed were wrong. I thought computers were powerful and smart, that all I needed to do was to give them a few guidelines and they would begin to act as rationally and sensitively as a human. But the truth was that they were incompetent, extremely incompetent, in terms of intelligence. When I submitted my PhD thesis, I found that I was more confused than I had been three years before. I thought about quitting but a voice at the back of my mind said that I could not leave with nothing but confusion. So I stayed on.

J:  Did you ever regret your decision?

Zhang: Over the next four years, from 1997 to 2000, our team struggled to survive. It was an entirely new discipline and we needed to start everything from scratch, such as compiling the dictionary, a rather sophisticated and yet labour-intensive job. It could not generate any economic benefits, because it was still at the stage of fundamental research, so when competing with other, more mature fields in applied science, we were often the loser.

Also, because the discipline was new, many established experts in artificial intelligence found it difficult to understand. It turned out to be a huge disadvantage when we submitted our papers to leading journals, as editors and peer reviewers would treat them with suspicion and require us to submit more supplementary materials than usual.

The situation has changed a lot since the Chinese Academy of Sciences launched a project to fund creative ideas and new disciplines. Now we are working for several national projects on artificial intelligence and rush to meet our deadlines every day.

J:  What is your vision of HNC in the future?

Zhang: HNC can provide a single platform for a computer to understand different languages. My dream machine is a single computer that can read all the books humans have ever written and absorb every bit of knowledge that has been accumulated in history. When you want to know about something all you have to do is to sit in front of it and speak naturally about what you are looking for. You don't need to be precise, you don't need to be complete and you don't even need to know exactly what you are after. The computer can understand your needs better than you do because it knows more, much more than anyone else. This is the artificial intelligence that I am looking forward to

Appendix: