A LEGO toy with artificial intelligence
Creating artificial intelligence that would have a potential comparable to that of the human brain is a challenge facing thousands of scientists and IT experts all over the world today. An important step in this direction was taken by the researchers at the Institute of Biology and Biomedicine (IBBM) of Lobachevsky University: not only did they create a model of a spiking neural network that simulates the work of a living brain, but they also constructed a self-learning robot based on this model. Sergey Lobov, Senior Lecturer at the IBBM Department of Neurotechnology, tells the correspondent of the “Rossiyskaya Gazeta” about their project called "Spiking Neural Network with Synaptic Plasticity and Competition."
Neuroplasticity was predicted by Descartes as early as almost four centuries ago, of course, in terms of that time. In the 20th century, Donald Hebb proposed the modern concept of synaptic plasticity: if a neuron is constantly involved in the excitation of another one, then the connection between these neurons is strengthened. Since the experimental confirmation of synaptic plasticity, it was assumed that it is the basis of learning and memory, but so far no detailed models or robotic systems have been proposed that could be trained using this principle. It is also known from physiology that there exists the phenomenon of competition between neurons in the form of lateral inhibition.
As Sergey Lobov explains, "if there is no competition between neurons, then there is no training." The robotic toy car, assembled using a LEGO kit, functions like this: at first it moves and runs into an obstacle, and a diode lights up to signal that “it hurts”. The robot remembers this "pain." On the next attempt, the robot avoids the obstacle and travels around it. This behavior is not programmed, it is a newly acquired skill. It is similar to the conditioned response in Pavlov’s dog. Moreover, the robot can be retrained (which is even more difficult than starting from scratch): even if its ultrasonic sensors are swapped, then over time it will still learn to go around the obstacle. If we compare this robot with an unmanned vehicle, it turns out that it definitely has a lead. To train the “mind” of an unmanned vehicle, one has to use a set of actions that are predefined in a huge table. In this case, a lot of resources and mostly manual labour is required to create such a table - for example, it should be clearly defined what is the actual road, what are the obstacles, etc. The Nizhny Novgorod robot has a self-learning capability without any prompting.