"Thinking like a baby", Deep Mind's latest study makes AI the same intuitive perception as intuitive perception

Artificial intelligence will play chess, and even defeat the world Go champion, which is old -fashioned.

Don't look at AI so powerful, it just uses the method of deep learning. Through countless training, it can be used to find the optimal solution from the training database. Think about itself.

On July 11, Deep Mind's latest article about AI has made breakthrough progress. Researchers use the knowledge in the field of development psychology to allow AI to "learn to think" like human babies. Earlier, the most advanced AI system was still difficult to capture the "common sense" knowledge in daily human scenes, such as guiding prediction, reasoning and action.

It can be said that the research of Deep Mind has greatly promoted the development of artificial intelligence in intuitive perception. Since then, artificial intelligence is really more like "people".

Written article | Xu Shilu

As early as when the robot came out, many people imagined: If the robot can think like a person, what will the world look like? (I believe that the corresponding movie scenes have emerged in your mind)

This issue has been sealed for many years, but a recent study published in "Nature · Human Behavior" shows that artificial intelligence may have the ability to "think like a baby": it can understand intuitive physics like a baby. It is worth mentioning that this is the result of the results of the well -known artificial intelligence company Deep Mind.

Luis Piloto and his colleagues made a deep learning system that can learn intuitive physics called Plato. The system contained in Plato is inspired by the study of infant learning methods. In addition, Plato follows the theory that objects plays a core role in the representation and predictions of our surrounding physical world.

Specifically, the researchers train it by watching many videos depicting simple scenes to Plato. For example, the ball falls to the ground, the ball rolls behind other objects again, and many balls play between the balls. After training, Plato showed "surprise" as human babies when seeing the meaningless scene (such as the objects passing through each other but not interacting with each other).

Surprisingly, Plato only watched the 28 -hour video and gained the above learning effect.

We can't help asking, what is intuitive physics, and how do babies understand it? How did AI learn this?

What is the intuitive physics in the eyes of the baby?

First of all, let's clarify a concept, what is intuitive physics? We can simply understand it as "intuition" or "common sense".

For example, we left a string of keys above the table, and everyone knew that the key would not float in the air, and we would not pass through the desktop to the ground, but it would fall on the desktop.

This is "intuitive physics". It is the basic physics concept we understand the world, and it is also a key component of "common sense" in thinking.

In the field of development psychology, intuitive physics is divided into 5 concepts:

1. Continuousness: The object will not be transmitted from one place to another, but there is a certain continuous path in time and space;

2. Object persistence: Objects will not disappear when they are invisible;

3. Solid: Objects will not penetrate each other;

4. Unknown: The attributes of the object (such as shape) will not be changed;

5. Treating inertia: The path of object movement is consistent with the principle of inertia.

Is it good to understand every one? That's right, these are the concepts of "common sense" that we easily understand and accept.

If the drop of the key is against our common sense, such as suspended in the air, or passing through the table, or playing from the desktop DUANG, it is old, or even turns into a liquid, then things will exceed us. The expectations became weird.

Faced with such weird events, everyone would be surprised. Even if the three -month -old baby is the same, they will be surprised to such a phenomenon of violation of intuitive physics. This surprise reaction is called violation of expectations (VOE) effects.

As for whether the baby's understanding of the world is the same as adults, there is a controversy about "congenital" and "acquired". Many development scientists believe that this is "congenital", and some scholars also support "acquired" from scratch to some. theory.

Can we design a process of simulated baby thinking, and through appropriate training, can AI think like a baby?

How to train AI like a baby?

In order to explore the "congenital" and "acquired" issues that are endless, Piloto et al. Using the PLATO simulation system to test whether the deep learning system can learn from the visual animation to obtain an understanding of intuitive physics.

If the theory of "acquired" is correct, the key to intellectual development is to exercise a lot of experience and a lot of data.

Plato simulation system consists of two modules: feedback perception module (left) and cyclic dynamic predictor module (right). In the feeding perception module, the image is converted into a set of object code by the encoder, and the object code is decoded into an image image through the decoder module. Use the differences between reconstruction and original images to train parameters of the encoder and decoder. In the cyclic dynamic module, the dynamic module predicts the object state of the next moment through a pre -trained decoder. The training data used by this study is less amazing, so that AI has a stable violation of expectations. In the face of the situation that exceeds the concept of intuitive physics, the Plato simulation system will show the VOE effect of robust (that is, stability). This can explain the concepts of some intuitive physics, but this effect is not exactly the same as the phenomenon seen on the baby.

Studies have found that although the experience of visual animation has contributed to the development of intelligence, it is not enough to explain the phenomenon we see in babies.

In other words, the development of intelligence does not depend on the experience and data of "acquired". To form a complete intelligence, some innate awareness is needed. This study has established an interesting intermediate position between the "congenital" faction and the "acquired" faction.

When video becomes impossible in intuitive physics, AI's surprise will increase significantly.

In addition, the Plato simulation system can summarize the expectations of intuitive physics as a set of objects and events that are different from training. Secondly, although it only receives relatively few visual animation training, this simulation system can also successfully demonstrate and learn, and there are similar features in infant research.

It can be seen that the researchers have unexpected results and conclusions that integrate the work and development of simulation and development psychology.

At present, the research team is expanding the research of violations of the expectations to the field of neurophantics. This may open up new possibilities for subsequent research and provide more possibilities for the development of AI.

Reference

1. Can a computer Think like a baby? Nature news.

2. Intuitive Physics Learning in A Deep-Learning Model Inspired by Developmental Psychology.https: //doi.org/1038/s41562-01394-8.

Source: Return to Park Popular Science China-Creative Cultivation Plan

- END -