It can teach a robot to perform complex tasks, typical of a trained human being. However, when it comes to imitate babies, things change. Yes, as strange as it seems. And it is that human beings in our first months of life have a vision of the world that has not yet been manipulated. Our learning, although logically influenced by upbringing, has a high dose of intuition, which is clearly seen when a baby is surprised to see movements that do not obey the laws of physics. This is something that she has not been able to do until now. artificial intelligence (AI). But only until now, as a team of scientists from the company deep mind has designed an algorithm that can be trained in a similar way to a baby of a few months.
During its training, the artificial intelligence algorithm, called DISH for Physics Learning through Machine Coding and Object Tracking, he was exposed to videos of balls moving in different conformations. Thus, he was able to gradually acquire the knowledge that allowed him to discern the basic laws that can differentiate a newborn baby.
Like a month-old child, the AI does not understand physical depths as such. However, it can differentiate between something possible and something impossible. Just as the baby is surprised by something that cannot be, artificial intelligence detects that there is something wrong. After all, it is his way of being surprised.
The basic laws that a baby masters without realizing it
These scientists drew on the studies psychologists have done over the years on the babies behavior. It is known that these are capable of distinguishing the operation of a series of concepts.
First of all, the solidity. That is, they detect that two solid objects cannot pass through each other. On the other hand, they are fully aware of the permanence. This means that an object that is fixed in one place cannot suddenly disappear. Finally, they detect continuity, so an object moves through space following a coherence. You cannot suddenly jump from one place to another.
This is the basic thing that babies do. However, these scientists added to their artificial intelligence the ability to discern the immutabilitywhereby the properties of objects cannot change, and the motion inertia.
All of this was taught to PLATO through a series of videos with balls. These were part of the training process, by which data is entered into an artificial intelligence algorithm so that it can create patterns that allow it to carry out a certain action. For example, if you want to train an AI to detect breast cancer in mammography images, it is previously trained with thousands of photographs with and without cancer, so that it establishes its own criteria.
It must be said that, in part, this is also what babies do. They observe the world and draw their own conclusions. Although it seems that they are better at it than robots. Incredibly.
The successful learning of PLATO artificial intelligence
According to the authors of the research in a study published in Nature Human Behaviorthe artificial intelligence training was so effective that in some cases could be carried out in just 28 hours.
After that time, if PLATO was exposed to new scenes, this time real, he would surprised with which they didn’t make sense. I even detected something strange in physically impossible scenes that did not appear in any of the videos with which he was trained. But does this mean that we can now speak of an artificial intelligence capable of behaving like a baby? Not so fast.
These scientists consider that PLATO is still far from the capabilities of a three-month-old child. And it is that, although it made great advances, it still behaved with some limitations. For example, his surprise was much less in implausible scenarios in which there were no objects.
In addition, the scientists had to provide him with some additional information so that he would acquire the knowledge that he later showed. This could answer a question scientists have long asked about the human species. To what extent the breeding exceeds the innate?
In short, this artificial intelligence is a step forward in the race to obtain robots that are increasingly similar to people. But it is also a new method to understand humans. Because, although science makes every effort to achieve it, it is clear that understanding us is not easy at all.