Head of the Department of Neurotechnologies of the Institute of Biology and Biomedicine of UNN Viktor Kazantsev
- Viktor Borisovich, recently scientists from the American University of Johns Hopkins proposed to create a biological computer, the main of which should be organoids - artificially grown cells of the human brain. Please tell us what a biocomputer is? What are the main concepts of such systems being developed in the world and in Russia?
- Let's start with the terminology: the very word "biocomputer", which is used by the media, is not entirely correct. In fact, the right term for this area of research has not yet been found. We can call such constructions neurotechnological devices.
Biological cells, including neurons, have a completely different nature than a computer. They do not work on the principle of binary logic, neurons do not operate with bits, they produce electrical and chemical signals - analog, not digital.
We're so used to the digital environment that it feels like numbers have always been with a person, but that's not the case. The human mind created mathematics quite recently by the standards of evolution.
Returning to the question, I will say that research with live neurons is now being conducted very actively, both in the world and in Russia – in particular, at UNN. We work with both animal brain cells and human neurons. Cells are planted in a test tube, where an optimal nutrient medium is maintained for them. An important difference between neurons and other cells is that they do not divide – this complicates our work. However, they can grow and establish new connections with each other through synapses (areas of the neuron responsible for contacts with other nerve cells. – RT). As a result, we get a test-tube model of a part of the brain that can live for several months.
In practice, such cell cultures are now actively used in the development of pharmaceuticals. This is very convenient - you can immediately see how a particular chemical substance affects the work of neurons.
- © Viaframe
As for the use of such cultures for neurotechnology, although neurons are not designed for computation, you can try to teach them this. We conducted such experiments, we managed to teach neurons to remotely control the robot via the Internet. It was a long time ago, about 10 years ago, when it was written about in the media.
But, of course, it is difficult to talk about serious computing systems in this case, because in a test tube neurons grow chaotically, in all directions. While in the brain, neurons are organized into clear structures.
However, now the world is working on the formation of a given cellular architecture in a test tube with the help of a network of microchannels etched on the substrate - microfluidic channels. And if it is possible to structure the cell culture, to form the input and output layers, then this already resembles those formal, mathematical algorithms that are used in computer neural networks. Only in this case, the work will be performed not by logical elements - ones and zeros, but by living cells.
— You described one of the concepts of a conditional biocomputer or neurotechnological device. And what other directions are there?
- Now three main approaches are developing. I have already mentioned the first one – the use of living neurons to perform some tasks in conjunction with technical devices.
The second way is when you understand exactly how the brain works and create a mathematical model that can reproduce its individual functions.
The third way is when, based on this model, you create a physical model and design a technological device, for example, a chip that can work as some part of the brain works. But without the participation of living neurons.
This path can be compared to the famous movie "Terminator", where the robot received computer intelligence similar to human. So far, people have not learned how to make such systems – primarily because science does not know exactly how the brain processes information. Much is known, but there is no complete picture.
That is why there was an interest in the use of living neurons as elements of a computing system – we do not know exactly how they work, but we can try to use them. Including to understand how our brain works.
However, you can't just take and connect neurons to a regular computer, because neurons exchange analog rather than digital signals, as I said earlier. Therefore, we also need an input and output system, an interface. One of the variants of such a system is the construction of memristors. An analogy for such neurohybrid systems can serve as another film - "RoboCop", where the human brain was placed in a robot.
- © 20th Cenury Fox
There is also a third direction of research, here as an analogy I will cite the film "Avatar", in which the inhabitants of the planet Pandora "connected" to dragons and controlled them. Approximately this principle of operation in the so-called human-machine interfaces. They are usually used in medical rehabilitation, when a person with motor dysfunction can control machines directly – brain signals that are read by special sensors. This is already biometric monitoring, which is used in medicine.
— What are the advantages and disadvantages of using biocomputers compared to traditional computers?
- As I said at the beginning, the human brain throughout the history of its formation has not faced tasks of a computational nature. It doesn't make any sense for us to try to compete with a car in this.
However, the brain solves other tasks incommensurably more efficiently than the most powerful computers. I will give a fairly simple example - the motor skills of body movements, limbs. If we try to create a computer that could control the movement of the hand as accurately and synchronously as the brain does, we will need simply colossal computing and energy resources. Hence such interest in neurotechnologies, and the prospects for their commercial use. If it were possible to create such a controller that could control the same number of drives with the same accuracy as the brain controls muscles, then no semiconductor processors could compare with it.
— And what are the prospects for molecular computers that perform calculations using a sequence of molecules in DNA?
- This is a slightly different scientific field. Frankly, I am skeptical about this direction, because DNA is a microstructure where everything is arranged on the basis of statistics, and not clear patterns. Such studies can be useful to geneticists and molecular biophysicists, biochemists, as well as for solving bioinformatic problems. However, it is unlikely that such calculations will be able to find application in the real world.
- © Tom Williams/CQ-Roll Call, Inc
Head of the Laboratory of Memristor Nanoelectronics of the Scientific and Educational Center "Physics of Solid-State Nanostructures" of UNN Alexey Mikhailov
— Alexey Nikolaevich, earlier you and your colleagues implemented a project to pair electrical circuits with a neural network of brain cell culture. Please tell us more about this.
"It was a relatively simple system — we connected an artificial neural network based on memristors to a living system of nerve cells so that they worked together and exchanged signals. The work was carried out within the framework of the project of the Russian Science Foundation (RSF). The use of living neurons is associated with a number of problems - living cell cultures are constantly changing, aging, and as biological degradation changes, the functional connections organized in ordered cultures also change. The task of an artificial network of memristors was to track these changes. In this case, a special technique was invented to adapt the external stimulation of culture in order to minimize the consequences of these changes. The hybrid neurotechnological system could not only classify the response of living cells to external stimuli, but also worked in it and feedback that helped live cell culture adapt to its own changes.
- What are memristors?
Unlike traditional semiconductor devices, such as transistors, which are logical digital elements, memristors are analog elements. They have a very simple structure: it is a layer of oxide dielectric of nanometer thickness, placed between layers of conductive metals. Part of the oxide layer contains an excess of oxygen atoms, the other, on the contrary, is deficient, and is also able to conduct current. Under a certain electrical influence, oxygen ions pass into the neighboring layer, which changes the conductivity of the memristor sections - the changes persist until a reverse force is applied. Thus, the memristor stores information - its principle of operation is very similar to the principle of memorizing information by living neurons. After all, the synapses of the neuron also change their bandwidth depending on what signal passes through them.
The bottom line is also that information is not only stored, but also processed in one place – this happens in neurons, and this principle is implemented in memristors. Unlike a classic computer, where there is a separate processor, separate memory, and between them there is a constant exchange of data, which requires energy expenditure, the memristors themselves both store and process information.
- © MF3d
Metal-oxide-metal memristors can be combined with traditional integrated circuits. We are now developing such hybrid devices together with the Sedakov Research Institute of Electrical and Electronics (Nizhny Novgorod) and NIIME (Zelenograd). On the basis of such systems, it is possible to create accelerators of neuromorphic calculations - this is a new hardware of artificial intelligence, which is currently being developed as part of the scientific program of the National Center for Physics and Mathematics.
Now neuromorphic computing systems very conditionally reproduce the architecture of the nervous system. So far, they are able to repeat only very primitive operations - if compared with the capabilities of brain structures. Memristors provide much more opportunities for creating brain-like systems that will no longer reproduce the mechanisms of the brain.
- You are also now creating a prototype of an artificial hippocampus. At what stage is the study?
In certain structures of the brain, for example, the hippocampus, which is responsible for memory and orientation in space, special neurons are concentrated that can selectively respond to individual images. Scientists sometimes call such cognitively specialized neurons "grandmother's cells" – that is, cells that allow us to recognize our own grandmother, for example. These neurons quickly and accurately respond only to images of a particular object and ignore the rest.
- © SEBASTIAN KAULITZKI/SCIENCE PHOTO LIBRARY
The peculiarity of these neurons is that they have a very large number of dendritic processes, that is, the "inputs" of information - and the more of them, the more selectively the neuron is able to determine a specific image. Mathematically speaking, the work of an individual neuron is usually reduced to the operation of scalar multiplication of the vector of input signals by the vector of synaptic weights. If two such vectors turn out to be co-directional, then they give the maximum amount to which the neuron reacts when multiplying - this is the moment of consolidation of information, recognition. The same input vectors, which are orthogonal to the weight vector, do not give a response at all.
In multidimensional space, with a large number of "inputs" of information, this recognition is much more efficient, since two random vectors are very likely to be orthogonal. Mathematicians managed to explain the unique selectivity of "grandmother's cells" in mathematical formulas. And with the help of memristors, it is very easy to reproduce this principle "in iron" and implement a physical model of the hippocampus. This work is now being carried out by us within the framework of the current RSF project.
— What practical application will such a system find?
First, with its help, we will be able to better understand how our brain and its certain structures work. Secondly, it will be able to quickly recognize various patterns – and, unlike modern neural networks, will not require long training. Such a system will be able to remember the image necessary for recognition from the first time.
I think that in the future a lot of technical products and solutions will be based on these technologies. For example, it will be possible to speed up the work of neural networks and artificial intelligence, to increase their efficiency. Plus, such devices on memristors will be more compact and energy efficient than traditional silicon chips. Quite in the long term, such systems will be able to replace the structures of the human brain in neurological diseases. Although it is too early to talk about it - fundamental and exploratory scientific research is still underway.