Artificial intelligence is still far from human at an important point

Japan's "K" supercomputer, one of the world's four most powerful supercomputers, has taken 40 minutes to simulate a second of human brain activity. Why is artificial intelligence still so far from humans? The following article may partly explain that reason.

The robot still has a long way to go before it can overcome human intellect. It is because people are quite good at visualizing the future and planning - one that shows the distinction between people and other species. Meanwhile, we are far behind robots in the ability to race in computing.

According to Berkeley University computer scientist Stuart Russell, people are very good at making "at the abstract levels" , a feature called decision-making hierarchy . Meanwhile, computers currently do not have the ability to decentralize decision making and perform many things at the same time as humans.

People also have the ability to look into the future, plan and make decisions based on abstract ideas.

Imagine when we went to a seminar in another city, Mr. Russell told Quanta magazine. When you arrive in a new city, you will not know the steps needed to go from the airport to the streets, you will only follow the signs of the airport exit. Just like when you get there, you can't plan ahead about the body movements needed to order a taxi. In your mind, you only have the thought of "I need to call a taxi" and the body will move accordingly.

" The way people do this depends very much on the abstractions and the high-level actions they accumulate , " Mr. Russell said. "Basically, that's the way we live our lives. The future is always spread out with lots of details close to us over time, but there are big landmarks, forcing us to come up. commitments made, are very abstract actions, such as 'obtaining a doctorate' "having children".

In other words, we make assumptions about the future based on our knowledge of the past world.

How people see the future

First , the most important factor is our emotions . They help us to make quick decisions about things and situations. For example, when people see a bear coming out of the natural world, they will feel the fear of forcing them to leave the area immediately.

However, AI (artificial intelligence) does not have these innate emotions to help the computer make decisions. A robot, unless specifically programmed to recognize the bear as a threat, will not automatically say that it is a large animal that can seriously harm me.

Second , the human brain is capable of seeing an object that we have never met before, will use our current intelligence and insights about the world, to visualize whether this object will how. According to computer scientist at NICTA, Toby Walsh, that is called speculative reasoning.

Illustrations for Common Sense - knowledge from common insights.

For example, every chair that we see through is a bit different - in terms of chair backs, special details, materials or colors. But when we meet something flat, supported by four legs, with a back, we will know that it is a certain type of chair and can sit.

Imagine when you want to call a taxi, but since you have never called a taxi before, now you don't know how to call. Then you see a taxi coming to pick up a person who has raised his arm. By speculation according to common knowledge, you will guess that this is the way to call a taxi.

By doing so, it is because people can recognize things like the "cause-effect" principle and apply it to the decisions they make. But this kind of empirical judgment is based on pre-existing knowledge, which is difficult to program into the artificial intelligence system.

According to computer scientist at New York University, Mr. Ernest Davis, for computers, to apply reasoning reasoning - this result, they need to know exactly the characteristics and parameters of everything in the given situation. guess. Computers also need to see that context many times to fully understand what is happening in this context. People do not need that - we can use our ability to imagine the future to think about the different possibilities and the reasons why it happens.

Third : According to Thomas Dietterich, Chairman of the Association for the Advancement of Artificial Intelligence, another thing that makes people go beyond robots is the ability to do many things at once. Most of the artificial intelligence programs we have today are very good at a particular job, like playing games or playing chess. But people can do more than that, "like finance, sports, raising children, collaborating with each other or opening packages" all at once.

"There are no AI systems that can reach these capabilities, especially when combining these capabilities with the vision, language and body movements, " Dietterich said.

How we improve AI

So, how can we create an inferred AI and make decisions like humans? Dietterich said that finding a way to "describe knowledge and information" in AI has also been a big step.

Others, like Mr. Peter Norvig, think the biggest obstacle to getting smart robots like humans is conquering the sense of perception . "We are very good at collecting data and developing algorithms to reason with those data. But that inference is only equivalent to data, meaning that it is a step away from reality. " Mr. Norvig said in his email.

According to Mr. Norvig, AI systems will better infer when they can see and perceive the world around them . "I think the inference will improve as we develop systems that can constantly feel and interact with the world. This is in contrast to passive learning systems from information. that others choose for " . Mr. Norvig said. Learning how to look, hear and feel will require robots to learn as to how children learn to walk - through trial and error.

The biggest obstacle to getting smart robots like humans is conquering the sense of perception.

As you grow up, you can learn about the world in a number of different ways. You can have a parent or teacher show an object and say what they are called. The same thing is being done with machine learning algorithms, currently used to train the AI ​​system.

But there are also many other self-learning possibilities that are hidden, and based on our ability to use reasoning to fill the gap on our previous knowledge. This is the kind of learning that today's artificial intelligence programs are still lacking.

Machine learning systems are learning from the beginning of new tasks on a daily basis. But according to AI researcher at Google, Samy Bengio said, this is extremely time-consuming and the future smart machines need to be able to learn without this basic approach."We need to study more about continuous learning - that's the idea that we don't need to start retraining the models from scratch every time we have new data or algorithms to try. half". Bengio writes. "Having a lot of difficult tasks will certainly take a very long time to improve."