How it works?
The basic principle of AI-powered kindergarten is the interaction between machines and people. The machines are still dumb and inexperienced at first, but they constantly receive feedback, just as a child does from its parents. Adults play, interact with him, praise him when he does something right and correct him when he does something wrong. The same should be done with machines. What they learn in this way, they pass on to other robots.
Does this mean that an intelligent system cannot consist of only one machine?
I agree. Because he has to get feedback from other people and machines over and over again. The more these experiences are collected and shared at the same time, the more efficient the entire system will be in acquiring the necessary competencies. Let’s say 100 people each train a small robot. Then each hundred small AI trains a large intelligence together, which in turn passes its aggregate knowledge to each individual small robot. These then interact with the humans again and receive feedback from them, in order to eventually pass on their new experiences to the big robot. You have to go through this cycle over and over again. The process is somewhat reminiscent of what happens in our brain during sleep.
In which way?
One of the goals of our dreams is to connect the things we have learned during the day with the knowledge that exists. Overnight, channels and membrane receptors are modified in such a way as to compare new learning content with older content. This is the main challenge of AI research. When we train a traditional artificial neural network, we repeat the same example countless times. For example, if we want to teach him to recognize cars and flowers, we need to present the same number of pictures of cars and flowers in random order. This is the only way to balance the two representations in the system. If she had been trained with a million pictures of cars but only ten pictures of flowers, she would have mistaken all things for cars. On the other hand, for children, one flower is enough to understand the essence of the flower. Because their minds can connect what they have just learned with what they already know overnight.
The Chinese Room – What Machines (Can’t) Do
A computer that can answer questions or play chess is considered smart by many. But is this really comparable to the human mind? Experts have been debating this question for decades. Philosopher John Rogers Searle (1932) of the University of California at Berkeley was a skeptic from the start.
To demonstrate the possibility of computers thinking, he designed a famous thought experiment in 1980: a man in a closed room. Through a slit in the door, someone pushes him a piece of paper with a story written in Chinese on it. Because the man does not speak Chinese, he cannot decipher the characters and therefore cannot understand the story. A second paper comes through the hatch, this time with questions about the content of the story, again in Chinese characters. However, in the room there is a guide to the grammar of conversion, a kind of database of Chinese phrases. The man pulls the strings from the guide that match the questions on a piece of paper and takes them out of the room. A Chinese reads the sentences and comes to the conclusion that the man in the room understood the story. He just followed the set rules without understanding what it was all about.
According to Searle, this is consistent with the way a machine with (supposedly) artificial intelligence works. She can only apply grammar correctly without recognizing the meaning. Therefore, the philosopher described it as a system of “weak intelligence”. On the other hand, an artificial neural network with “strong intelligence” must understand semantics – just like humans.
How long will it take to attend a kindergarten machine powered by artificial intelligence to be as smart as a human?
It is impossible to predict this. When we raise a child, it takes about 20 years before they can easily take care of themselves. An artificial system cannot run faster. But it may also take 100 years.
Does speed also depend on computers?
Yes, this also plays a certain role. But the most important thing is to interact with real people and experiences. You cannot speed it up.
Is an AI-powered kindergarten just a theory yet, or have you already started implementing it?
So far it is just a concept. However, my colleagues and I have broken the comprehensive theoretical system into small sections and programmed the appropriate algorithms. We actually trained those – as proof of concept that the theory is headed in the right direction. We recently created a beta version that you can test at robotsgomental.com. However, there is a completely different problem: we have not yet found anyone who will fund the development.
Where can this algorithm be used?
We are currently developing specialized smart systems that can predict the course of events, such as stock prices. Later we want to extend their skills to include image and sound processing. Countless other applications conceivable of course.
In some science fiction scenarios, super-intelligent machines fight against humanity. Is this realistic?
No, the intelligence of artificial neural networks cannot grow uncontrollably. It would be like a perpetual motion machine: nothing comes out of nowhere. One can only become smarter by learning from the environment. An AI system will never be able to predict what knowledge it will acquire in the future. It also means: understanding of the world is limited; This applies to us as well as to machines. High intelligence is only possible through a very slow learning process.
The conversation was conducted by Anna von Hopfgarten, a Ph.D. biologist and editor at Gehirn & Geist.
Danko Nikolic studied Psychology and Engineering at the University of Zagreb in Croatia. In 1999 he received his Ph.D. in Psychology from the University of Oklahoma. In the same year he moved to the Max Planck Institute for Brain Research in Frankfurt am Main, where he began recording neural activity in different parts of the cat’s brain in parallel using so-called multichannel electrodes. In 2010 he was appointed Associate Professor at the University of Zagreb and in 2014 Associate Professor. Since 2016, he has been dealing with questions from the field of data science and developing economic applications for artificial intelligence systems.