Australian biologists from Monash University linked hundreds of thousands of neurons to a computer and used a reward system to make them learn coordinated actions. Such a "protomouse" easily mastered the computer game Pong, which requires hitting a ball with a virtual racket. The experiment is described in an article published in the journal Neuron.
Brett Kagan and his colleagues obtained biological neural networks "in vitro" using rodent and human stem cells. The system of about 800 thousand cells grew on an array of microelectrodes, which provided the exchange of signals with the computer, the scientists called it DishBrain, brain in a test tube. And the computer game Pong, a simple two-dimensional analogue of ping-pong in which you have to hit a virtual ball with a virtual racket, served as a test of DishBrain's ability to adapt and process sensory information efficiently. In other words, to learn.
The key to this was the feedback that the neurons received in the form of electrical signals generated by the specially developed SpikeStream program. It made it possible to encode the movements of the game ball: the electrical stimulation in one or another part of the DishBrain indicated the position of the ball in space, and its frequency indicated the distance to it. Similarly, the output signal was encoded: the localization of neuronal activity corresponded to the direction of the racket's movement, and its frequency - to its speed.
DishBrain is much simpler than even the most primitive brain; it has no dopamine or other reward system. This is why the principle of free energy played such a role, according to which living systems seek to minimize entropy, the uncertainty of their environment. An unpredictable stimulus is applied to cells, and the system as a whole reorganizes its activity so as to play the game better and minimize randomness, says Brett Kagan. One could say that by kicking the ball and getting a predictable response, it creates a more predictable environment for itself.
If DishBrain made a mistake in the game, it received chaotic electrical signals lasting several seconds in response. If the neurons kicked the virtual ball, however, the response was a brief and predictable signal. And this approach worked: it turned out that literally in five minutes the system learned to coordinate the activity of individual cells, adapting and successfully learning to play.