New frontiers for research: the robot that learns from real neurons arrives

It recently appeared in the journal Applied Physics Letters , an experiment that saw a robot capable of learning from real neurons. It is an extremely interesting work due to the extent of its repercussions as it changes the approach that has been used up to now in this field . In fact, until now, typical studies have seen computers learn in the field using artificial intelligence algorithms. In this case, the computer (and the robot which is the component capable of interfacing with reality) only acts as an intermediary towards a culture of real nerve cells .

Neural networks: what they are

To get to the experiment of the scholars of the University of Tokyo it is good to take a step back, to the 50s of the last century. At that time, there were the first attempts to recreate the mechanisms of human thought through the computer. A neural network is a mathematical model that tries to reproduce the functioning of a biological neural network . In particular, through a non-linear structure, the neural network receives inputs on numerous receptors which in turn forward their output to another series of nodes called neurons. And so on on various levels until the final output is produced.

Neural networks can, therefore, be trained to respond to various problems that, due to their complexity, cannot be solved with linear functions. Among the most typical applications are:

  • Statistical approximation or regression functions;
  • Classification problems, such as images, through typical classes and patterns;
  • Finally, data processing, including “filtering” (noise elimination), clustering, signal separation and compression.

Activities that belong, in fact, to the basic intellectual abilities of human beings and that we have tried to reproduce through these mechanisms.

The robot experiment interacting with real neurons

The great revolution that brings the experiment conducted by the researchers of the University of Tokyo is the presence of a real network of neurons instead of a series of artificial neuron levels . In this case the system is made up of real neuronal cells connected to a computer which, in turn, acts as an intermediary with the robot. In fact, the small robot, about the size of a hand, is left free to move on a surface containing obstacles.

The robot's goal is to reach the destination without knowing how to do it but through a mechanism called “trial and error” , or “ trial and error ”. At each wrong attempt by the robot (for example when it bumps into an obstacle), the cultured neurons are stimulated by the computer through an electrical impulse. In this way, the robot has learned to move between the obstacles of the path to arrive correctly at the final finish line.

Study co-author Hirokazu Takahashi talks about the ability of living neuronal systems to steal coherent output from a state of chaos , as the network used had connections between causal neurons. Certainly a great achievement, even if still in an embryonic state, but the hope of scholars is that if this approach proves to be sufficiently stable and safe, the cultivated neural networks could be implanted in the cases of damaged biological networks.

The article New frontiers for research: the robot that learns from real neurons arrives from Tech CuE | Close-up Engineering .