Google and Microsoft does not share stage often are
more ferocious competition in areas such as research on the Web, mobile
and cloud computing. But the rivals to agree on things like the importance of the artificial intelligence for the future of technology.
Peter
Norvig, Google's director of research, and Eric Horvitz, a scientist of
renown in Microsoft Research, recently spoke at a joint hearing in the
Computer History Museum in Palo Alto, California, with the promise of
AI. Then, the couple spoke with the editor review
of IT technology, Tom Simoniti what AI can do today and what I believe
will be able to tomorrow. Artificial intelligence is a complex issue and some answers have been edited for sake of brevity.
Technology
Review: The two spoke on stage for AI there has been progress in recent
years through the use of learning techniques that have large amounts of
data and understand things like how to translate the text or
transcribed speech. What happens with the areas that we want to help AI that there is no data much to be learned?
Peter Norvig: what we do is look under the street lamp of the keys fell because the light is there. We are very well with the text and speech, because there are many data in nature. Analysis
[break the elements of grammatical sentences] does not occur naturally,
perhaps someone from the duties of the language, we have learned that
without [tag] data. One of my colleagues is
trying to avoid it looking at what parts of the text were made links
online that may indicate that a particular part of a sentence is.
Eric
Horvitz: I have often thought that if had a service of cloud in the sky
that record every word and asks what happened then, all the
conversations in all Beijing taxis, for example, it may be possible AI
has to learn to do it all.
Seriously, if we can find a way to capture a large amount of data in a way that preserves privacy, we can make this possible.
Is
not difficult to use the machine if the formation of workout data is
not marked and explained, to give the AI a "truth" to boot from the?
Horvitz: You do not need to be fully labelled. An
area known as supervised semi-aprendizaje has shown that even if 1
percent or less of the data labels, you can use to understand the rest.
But the lack of labels is a challenge. One
solution is to actually pay the people of a small amount to help a
system with data that cannot be understood, so microtasks labeled as
screenshots or other things. I think that the calculation used to increase the human AI is a very rich area.
Another
possibility is to build systems that comprise the value of the
information, which means that you can automatically calculate what the
next question is better, or how get the most out of one tag additional
or element information provided by a human being.
Norvig: You don't have to tell an entire system of learning. There is a type of learning called learning reinforcement, which has just given a prize or punishment at the end of a task. For example, it has lost a game of checkers and you do not say that you have the wrong and have to learn how to get the reward that the next time.
All this is very different from the early days of the AI in the years 50s and 60s, when researchers have made bold predictions about the correspondence of the human rights and tried to use high level rules for the creation of the intelligence. Is your system of automatic learning working in the same high level of the rules themselves?
Horvitz: learning systems can lead to high standards of the situation for action, for example, take a set of [physiological] symptoms and test results and diagnosis of cough. But this is not the same as the General rules of intelligence.
It is possible to run the lowest we do today to respond to ideas from top to down bottom-up one day. The revolution that Peter and I were part of the IA was that decision making under uncertainty is so important and you can do with probabilistic methods. With the revolution of AI has perspective probabilistic: we are a very limited and incomplete personnel is inevitable.
Norvig: In the early days, it was logical that the game's artificial intelligence in hand, and the question was how to use it. The study has become the study of what these tools are good for, such as chess. But you can have the things that are true and false, and you can't do many things that we want to do, so we went to the probability. It took some time in the area to identify other areas such as probability and decision theory, they were there. The transfer of these two approaches is a challenge.
As we see the most direct evidence of AI in real life, for example, Siri, looks like a kind of design problem has been created. The people who create AI need that are acceptable to our own intelligence.
Norvig: You don't have to tell an entire system of learning. There is a type of learning called learning reinforcement, which has just given a prize or punishment at the end of a task. For example, it has lost a game of checkers and you do not say that you have the wrong and have to learn how to get the reward that the next time.
All this is very different from the early days of the AI in the years 50s and 60s, when researchers have made bold predictions about the correspondence of the human rights and tried to use high level rules for the creation of the intelligence. Is your system of automatic learning working in the same high level of the rules themselves?
Horvitz: learning systems can lead to high standards of the situation for action, for example, take a set of [physiological] symptoms and test results and diagnosis of cough. But this is not the same as the General rules of intelligence.
It is possible to run the lowest we do today to respond to ideas from top to down bottom-up one day. The revolution that Peter and I were part of the IA was that decision making under uncertainty is so important and you can do with probabilistic methods. With the revolution of AI has perspective probabilistic: we are a very limited and incomplete personnel is inevitable.
Norvig: In the early days, it was logical that the game's artificial intelligence in hand, and the question was how to use it. The study has become the study of what these tools are good for, such as chess. But you can have the things that are true and false, and you can't do many things that we want to do, so we went to the probability. It took some time in the area to identify other areas such as probability and decision theory, they were there. The transfer of these two approaches is a challenge.
As we see the most direct evidence of AI in real life, for example, Siri, looks like a kind of design problem has been created. The people who create AI need that are acceptable to our own intelligence.
Norvig: This is actually a set of problems at different levels. We know that the human visual system, making it different color keys can mean, for example. At
a higher level, the expectations in our minds about something and how
should behave based on what we believe is our way of thinking and your
relationship with us.
Horvitz: intersection AI is growing with the interaction field titled [the study of the psychology of how we use and think about computers]. The idea that we have more intelligent things that work closely with people really focuses on the need to develop new approaches to the intersection of human intelligence and artificial intelligence.
What can we learn more about AI to make it more compatible with human beings?
Horvitz: One thing that has been my research panel pushed to give computers is an understanding of the entire system of human attention, which is the best time to leave a person. This has been a subject of research between researchers and product teams.
Norvig: I think they want to understand the human body, much more, and a way of doing this can be seen in Microsoft Kinect. There is great potential for systems to understand our behavior and body language.
Is there an AI Kinect?
Horvitz: There are many machine learning at the center of it. I think that the idea that we have an advanced AI and develop a device of consumption are sold faster than any other time in history says a lot about the field of AI. Learning machines also plays a central role in the Bing research, and can only assume also is important in the provision of Google search. So people search to your web use of AI in their daily lives.
One last question: can you tell me a demonstration of the latest technology IA that has impressed you?
Norvig: Read an article recently by someone in Google about to return to Stanford in unsupervised learning, an area where the curves of our improvement over time does not look as nice. It is doing some really good results but it seems that learning when he does not know anything in advance may be about to do much better.
Horvitz: I was very impressed by the apprentice of learning, where the system learns by example. It has many applications. Berkeley and Stanford, both groups argue that in reality: for example, learn to fly helicopters at the back a human expert.
Horvitz: intersection AI is growing with the interaction field titled [the study of the psychology of how we use and think about computers]. The idea that we have more intelligent things that work closely with people really focuses on the need to develop new approaches to the intersection of human intelligence and artificial intelligence.
What can we learn more about AI to make it more compatible with human beings?
Horvitz: One thing that has been my research panel pushed to give computers is an understanding of the entire system of human attention, which is the best time to leave a person. This has been a subject of research between researchers and product teams.
Norvig: I think they want to understand the human body, much more, and a way of doing this can be seen in Microsoft Kinect. There is great potential for systems to understand our behavior and body language.
Is there an AI Kinect?
Horvitz: There are many machine learning at the center of it. I think that the idea that we have an advanced AI and develop a device of consumption are sold faster than any other time in history says a lot about the field of AI. Learning machines also plays a central role in the Bing research, and can only assume also is important in the provision of Google search. So people search to your web use of AI in their daily lives.
One last question: can you tell me a demonstration of the latest technology IA that has impressed you?
Norvig: Read an article recently by someone in Google about to return to Stanford in unsupervised learning, an area where the curves of our improvement over time does not look as nice. It is doing some really good results but it seems that learning when he does not know anything in advance may be about to do much better.
Horvitz: I was very impressed by the apprentice of learning, where the system learns by example. It has many applications. Berkeley and Stanford, both groups argue that in reality: for example, learn to fly helicopters at the back a human expert.
0 comments:
Posting Komentar