Human brain cells in a Petri dish or artificial intelligence, which of the two learns faster? In an unprecedented experiment , researchers compared the learning ability of artificial intelligence with … brain cells placed in something like an electronic petri dish. Surprisingly, when they “asked” these two entities to learn how to play the famous video game Pong, they were surprised to find that it was brain cells connected to an electronic device that learned to play this game the fastest, and not AI — as one might expect.
Machine learning is a branch of artificial intelligence that has flourished in recent years, with demonstrations and real-world applications becoming more impressive than previous ones. As a result, for the average person, advanced AI today has learning abilities far superior to simple primitive brain cells maintained in the laboratory… But a new experiment shows that this is not really the case…
Researchers now call their creations the “cyborg brain,” and the reasons for this seem obvious. Of course, this is not the first time that researchers have studied organoids (primitive mini-organisms created in the laboratory) of the human brain, but, according to Brett Kagan, lead author of the study and chief researcher at Cortical Labs, this is the first time that a mini-brain has been able to fulfill a specific purpose.
Each of the mini-brains created by Kagan and his team contains from 800,000 to 1 million living brain cells. Simply put, it’s about the size of a cockroach’s brain. Some brains were made up of mouse cells taken from embryos, while others were made up of human brain cells derived from stem cells.
In order for the organoids to interact with the virtual environment, of course, an electronic device was required. To do this, the researchers have grown cells on arrays of microelectrodes that can both stimulate cells and read their activity. The resulting system was named “DishBrain”.
To simulate a simplified version (without an opponent) of the Pong game (a table tennis video game), scientists came up with a simple idea: the firing of electrodes to the left or right of the array tells the mini-brain whether the ball is to the left or right of it, and the frequency of signals indicates proximity.
Certain patterns of activity between neurons are interpreted as moving the racket to the left or right. The computer reacts to this action, and feedback through the electrodes allows the mini-brain to learn how to control the racket. “When they are connected to the game, they believe that they themselves are a racket,” one of the researchers sums up.
Fast but limited learning compared to AI
Although the human mini-brain has been shown to learn to play pong faster than AI, it is still much more limited in the long run. Thus, advanced AI always turns out to be much better at games than DishBrain. But the learning speed of the organoids remains very impressive: while the AI tested in the experiment required 5,000 cycles to master the game, the mini-brains took only 10-15 iterations. This corresponded to about 5 minutes of training.
A brain made from human cells is much better at playing pong than a brain made from mouse cells. But since the source of the cells is different, the team cannot yet be sure that this is due solely to their human nature. In the future, it would be interesting to play these two types of mini-brains against each other, or human organoids against AI (at least in the early stages of learning, until it became invincible…).
However, this success was achieved not only by mini-brains, but also by researchers. Indeed, as noted by other researchers who commented on this study, the authors brilliantly managed to make the neural network comprehend digital data and simultaneously act in this environment. Therefore, the closure of the “action-perception” circuit is not only an exceptional technical achievement, but also brings us closer to the creation of real synthetic brains. In other words, the “cyborg brain”.
Kagan and his colleagues’ approach to learning is based on a theory of brain functioning called the “free energy principle” developed by Carl Friston of University College London. The basic idea is that even the neurons in the Petri dish are trying to create an internal model of their external world. In other words, they try to predict what will happen based on the data they receive, and do not like to be surprised.
That is why the cells are “playing a game”. When they play, their actions become more predictable. If they don’t play, they get random aversive signals. What is really amazing about this system is the intelligent behavior that occurs in an uncontrolled way.
The long-term goal of Cortical Labs is to develop a cyborg brain, which, according to Kagan, can be more intelligent than computer systems. There is also a more direct application. For example, learning how neurons learn so quickly and efficiently can help improve machine learning by reducing the large amount of energy needed and the time needed to learn. According to the researchers, the next generation of artificial intelligence should strive for the functional and thermodynamic efficiency of the biological brain. This work is a significant and perhaps even historic step in this direction.
Another potential application of these organoids is drug screening. According to Kagan, the introduction of experimental drugs into the mini-brain during the game can give a better idea of the effect of these drugs on the human brain than studying neurons individually.