How to break the boundary between four types of artificial intelligence?

The consensus among artificial intelligence (AI) is that intelligent machines with self-awareness are about to emerge. These machines do better than humans in terms of command understanding, picture differentiation, car driving and games. How long will it take for such a smart machine to enter our lives?

The White House’s new report on artificial intelligence expressed skepticism about this dream. The report said that in the next 20 years, it may not be seen that "the machine exhibits a wide range of intelligence that is comparable or surpassing humans," although the report goes on to say that in the next few years, "the machine will be in more and more tasks." To reach or even surpass human levels." But the report's hypothesis about how machines acquire such intelligence lacks some important points.

As an AI researcher, I am honored that my research area was mentioned at the highest level meeting of the US government, but the content of the report is almost entirely focused on what I call "boring AI." The report was rejected in half a sentence. The AI ​​branch I studied, in turn, explains how evolution can help develop an ever-improving AI system and how computational models can help us understand the evolution of human intelligence.

The report focuses on so-called mainstream AI tools: machine learning and deep learning. These artificial intelligence technologies have been able to excel in the "Jeopardy!" game and defeated the human Go master in the most complex games. Current intelligent systems can quickly process large-scale data and complex calculations. But they lack a single element, which is the key to developing a self-aware smart machine in the future.

What we need to do is not only to teach the machine, but also to break the boundaries between the four types of artificial intelligence and to overcome the obstacles between machines and humans.

The first type of AI: reactive machine

The most basic AI system is completely reactive and can neither form memory nor use past experience to guide current decisions. IBM's chess supercomputer, DeepBlue, defeated international master Garry Kasparov in the late 1990s, and it was this reactive machine.

DeepBlue recognizes the pieces on the board and knows how each piece moves. It can predict how the next step and how the opponent moves, and then choose the best mobile solution.

But it doesn't have any past concepts, and there is no memory of what happened before. Except for a specific rule of chess that is rarely used, it is not allowed to repeat the same movement three times. DeepBlue does not consider anything that happened before. It only considers the position of the pieces on the current board and then chooses one of all possible next steps.

This type of intelligence involves the computer's direct perception of the world and responds accordingly, rather than relying on internal concepts of the world. In a groundbreaking paper, AI researcher Rodney Brooks believes that we should only develop such machines. The main reason for RodneyBrooks is that humans are not good at constructing a precise simulation world suitable for computers, the so-called "representation" of the real world in AI academia.

The machine intelligence we are currently amazed at either has no such concept for the world or has limited concepts for specific tasks. The innovation of DeepBlue is that it is not a moveable solution to increase the next step. Instead, developers have found a way to reduce the number of mobile solutions, based on the results of each scenario, and to abandon exploring some potentially mobile solutions. Without this capability, DeepBlue will need to be stronger to beat Kasparov.

Similarly, Google's AlphaGo, although defeating the top master of Go, is not able to evaluate all potential mobile solutions. AlphaGo's analysis method is more complex than DeepBlue, which uses neural networks to assess game development.

These methods do make the AI ​​system perform better in a particular game, but they don't apply to other situations. These computer minds don't have the concept of a broader world—meaning that they can't perform tasks other than specific tasks, and they're easy to be fooled.

AI systems are not able to participate interactively in the real world. Instead, these machines are treated the same way each time they encounter the same situation. This ensures the reliability of the AI ​​system: I want my autonomous car to be reliable. But if you want the system to truly touch the world and react to it, it's bad. These simplest AI systems are never boring, interesting, or pessimistic.

The second category: limited memory

The second type of AI machine can observe the past. Self-driving cars have done something, for example, they look at the speed and direction of other vehicles. Observing the past can't be done in a short time, but you need to identify specific objects and keep monitoring.

These objects to be observed are added to the "representation" of the pre-programmed world of self-driving cars. The “representation” of the world also includes lane markings, traffic lights and other important elements such as road curves. These objects are taken into account when the driverless car decides to change lanes in order to avoid blocking other drivers or colliding with other cars.

But these simple pieces of information about the past are short-lived. Unlike the way in which drivers like to accumulate years of driving experience, they are not saved as empirical library information from which they can learn.

So how do you build an AI system that has a complete "representation" of the world, remembering historical experience and learning how to deal with new situations? Brooks is right, it's really hard to do this. Inspired by Darwin's theory of evolution, my research method can make up for the shortcomings of human beings to let the machine establish its own "representation" of the world.

The third type of AI: theory of mind

We may stop here and use this as an important demarcation point for current AI machines and future AI machines. However, we'd better discuss the types of "representations" of the world that machines need to build, and what they need to care about.

The more advanced type of AI machine not only forms a "representation" about the world, but also forms a "representation" about other media or entities. In psychology, this is called “theory of mind”—people, creatures, and other objects have thoughts and emotions that affect their behavior.

This is crucial for how humans form society because they allow humans to interact socially. If you don't understand each other's motives and intentions, or don't consider others' perceptions of yourself or your surroundings, working together is the most difficult, and you can't even work together.

If AI systems really enter our lives, they must be able to understand that each of us has our own ideas, emotions, and how we expect to treat ourselves. Then they must adjust their behavior accordingly.

The fourth type of AI: self-awareness

The final step in the development of AI is that the constructed AI system can form a self-representation. In the end, our AI researchers need to not only understand the consciousness, but also build a machine with consciousness.

In a sense, this is an extension of the "mental theory" of the third type of artificial intelligence. To a certain extent, consciousness can also be called "self-consciousness." ("I want that thing" is very different from "I know that I want that thing.") A conscious person knows himself, knows his inner state, and can estimate the feelings of others. When we think of traffic jams, the people who honk behind us are angry or impatient, because this is how we feel when we honk behind others. Without some kind of theory of mind, we cannot make these inferences.

While it is not yet possible to create self-aware AI machines, we should focus on memory understanding, learning, and decision-making capabilities based on past experience. This is an important step in understanding human intelligence itself. This is also crucial if you want to design or retrofit a special machine that is very good at classifying what you see.