ScienceHub # 07: Neurointelligence and Neuromorphic Systems
Not so long ago, we went to the neurophysiologist Mikhail Burtsev, head of the laboratory of neurointelligence and neuromorphic systems at the Kurchatov Institute. He told us what his employees are doing, what neuronal cultures are and why they actually need to study the brain.
To get to the institute's territory, you need to try hard at any laboratory: negotiations with the press service, an exact list of equipment (don't let Gd have an extra laptop) and a mandatory dress code - white coats, outstanding at the entrance. But it was worth it, and this is why: in general, PostScience has a long relationship with Mikhail Burtsev, he often comes to the editor to write down new videos, but it was great to communicate with him in his natural habitat - all scientists are right in the laboratories and show simple at first glance, but amazing things.
')
Burtsev Laboratory is engaged in the construction of neurointelligence systems. That is, trying to solve the complex problem of building artificial human intelligence, breaking it up into stages. First you need to understand how simpler networks work. Then, having understood the principles of how cells interact with each other, scale them at the level of the animal's intelligence.
Mikhail Burtsev: “It is important that we try to combine biological experiments and theoretical work in the field of neurointelligence systems in one laboratory. That is, we build computer models. There is some agent living in an intellectual environment. In order for this agent to solve some tasks in the environment that surrounds him, he must perform certain sequences of actions, receiving a reward for this. The agent needs a neural network that will control his actions.
The neural network will see what has happened in the environment. She must learn to solve problems in this environment. We have theoretical assumptions about how this should work. On the other hand, in order to test theoretical positions and, possibly, to extract some principles from real systems, we want to investigate what is happening in real networks in living neurons when they are being trained. Today came an important moment in the field of experimental neurobiology. A critical mass of technologies has been accumulated, which makes it possible to investigate neural networks alive in great detail. ”
The construction of neuro-intellectual systems as a direction in science has arisen within the research on the creation of artificial intelligence. It is due to the fact that in research on artificial intelligence there are two alternative approaches to how to create this artificial intelligence. One approach comes from psychology and our understanding of how a person thinks and thinks. In the laboratory, Burtsev is trying to reproduce in a computer program those intellectual abilities of a person that can be learned from psychology. The second approach comes from below. Since it is clear that our intellect is based on the interaction and collaboration of many billions of cells in the brain, we can try to model these nerve cells and build artificial neural networks of them that will perform intellectual tasks.
Both approaches have their own advantages and disadvantages. For example, the first approach comes from cognitive sciences. It allows you to quickly make applications for tasks that are associated with human speech activity, with the analysis of texts, to make systems that will communicate with a person in a natural language. But there is a problem in it: human consciousness is the most complex object of all that exists, and building a normal model of such an object is not easy, because it is difficult to understand which principles are necessary. And to make a complete normal model that would have at least the child’s intelligence is almost impossible now.
On the other hand, if you simply assemble a human brain from neurons, it will not work either and will not start thinking like a human. Knowledge of neuroscience today is not enough to build such a complex system.
You can put electrodes on your head, but pulling out information about how the brain works is very difficult. As studies of the last decades show, it is the intellectual component (brain training and memory) that is associated with the work of individual specific cells. When electrical activity is removed from the surface of the skull, a huge number of cells are summed up, so it is impossible to understand how the specific activation of cells in the brain has changed in specific tasks.
Another way to help see real cells in the brain is to implant an electrode in the head. So you can register individual cells.
MB: “There were a lot of experiments when it was shown that neurons can be found that are active only when a person sees certain photos. Or when he remembers something. Electrodes were implanted to the person in order to conduct further surgical intervention to stop epileptic seizures. For this you need to investigate in which place you need to cut a piece of the brain. Having such electrodes, we can check on a person how his brain works. My favorite video is when a person is shown cutting different segments of video clips. There, and the NBA, the clip of Madonna, a cartoon about the Simpsons, a picture of how the car enters Hollywood. We see in the video that the person is watching the video, and the Simpsons are shown to him, he begins to activate one cell. We hear knocking, cell discharges. In the second part of the video, a person is asked to tell what he saw. The man begins to say: "I saw the sign of Hollywood" and something else. Then he pauses. We hear this cell begin to activate. He laughs and says: "I also saw the Simpsons." This means that this neuron works both at the moment of the perception of this video and when it remembers it. He loses in his head how he sees this video. It is clear that we can learn a lot from these experiments. On the other hand, in order to understand how this all works at the level of the organism and of purposeful behavior, much more needs to be done. Imagine there is a composer who composes a symphony. The mystery is how a change in the behavior of individual cells in the brain leads to the fact that he comes up with a symphony that will appeal to hundreds of thousands of people? That is, how the change of thousands of cells at the same time led to the fact that he wrote some great work? How do these cells even know what this composer wants? What do these cells see?
These cells see some chemical signals that other cells send them. That's all. How to decode these signals? Do they come together to create this symphony? Or how to invent something or prove some theorem? This is the mystery over which we are fighting, trying to understand how this is happening. ”
To do this, you need to understand how several cells work at once simultaneously. Ideally, you need to have a brain in which you can see all the cells and then you can turn to neuronal cultures.
MB: “Neuronal culture is the cells that we took from the brain and put in a cup (in a nutrient medium). They grow there. And we know that they are starting to form networks there and show spontaneous electrical activity. We use the following cups in our experimental work. This is a round chamber in which we plant cells. There is also a nutrient medium. And along the edges of the square matrix there are contacts, to which electrodes are located, located in the center of the cup. The bottom of the cup is covered with a special solution that makes the cells adhere to the bottom. Thus, cells do not float in the nutrient medium. If this coverage is not present, the cells will begin to form clusters, because they have such a genetic need. They stick together and form the so-called neurospheres that float there. And to avoid this, we cover the bottom. And the cells stick to the bottom. And if they float, we cannot measure their electrical activity, because we cannot lead the electrode.
And so the cells are spread by a monolayer, or in two - three cells. And then all these cells are visible under the microscope. We will be able to land 5 -10 thousand cells there. Although, of course, in such a number of cells is difficult to understand. But still, comparing with the millions of cells in the brain, it is much easier. Accordingly, we can register what happens to these cells in terms of their electrical activity. This is important because the cells come together due to electrical activity. With the help of electricity, they emit chemical messages to each other. And when the cell is activated, we know that one cell passed something to another.
Obviously, if we want to build an intellectual system and understand how it solves some problem, then we will not achieve anything by simply placing the cells in a cup. There is some kind of activity, and no more. If you look in a microscope at the cells planted in a cup, they all lie first in separate circles. These are dissociated cultures.
A dissociated culture, when we take cells out of the brain, sever ties between them and plant them to the surface. They first grow on the surface. And after two or three days, they release processes and begin to connect with each other, feel each other and make contacts. When they have already established these contacts, they send messages to each other through these contacts. We can say that they have a fear of being alone. And they try to talk to each other. It turns out a social network.
The problem is that we do not understand what is happening there. United cells live some kind of life. We do not know what they think about when they are united in the brain. But we want to understand how neural networks adapt to something. To understand this, we need to set a task for the neural network that it will solve. Then we will be the result of solving this problem. And we will understand that in this moment the network has been trained, but not yet in this. ”
Burtsev, like many others, has a disciplinary area, and it also combines biology, mathematics and physics. More precisely, even people from Computer Science who can think abstractly, biologists who can do something with their hands. Often people move from one area to another.
MB: “For example, if I do machine learning, I know what an artificial neural network is. I can program them, build a program that will recognize images. But, if I want to develop these algorithms, to see where they can move on, I need to know how real neurons work. In general, feel what it is.
The experience that I get when working with real neuronal cultures and with real neural networks can be invaluable. In my experience, there are more specialists from the theoretical side in the study of neuro-intellectual systems among physicists and specialists from Computer Science. It would seem that if we have an interdisciplinary audience, then there should be an opposition between biologists, on the one hand, and physicists and mathematicians, on the other. In fact, it does not exist as such. Those biologists who are engaged in experiments need methods for analyzing and processing data. But they may lack skills from an instrumental point of view. Here they turn for help to physicists and mathematicians who understand how this data can be better processed and presented.
That is, physicists and mathematicians in this case do not provide conceptual, but technical assistance. And vice versa. When a physicist tries to do an experiment with cells, it is obvious that he needs help, telling what to do and helping to do this experiment. That is, there is mutually beneficial cooperation. ”
The main application of the study of neurointelligence systems are pharmacology and medicine. The second is an attempt to build intelligent systems. The last direction is the activity of the laboratory Burtsev.
MB: “The first layer of this activity is to teach the cells to control a real robot. We will launch a robot under water, collect volcano or debris from the street, solve the migration problem. But it is not yet clear how to implement this from the point of view of technology. First, we need to maintain sterile conditions all the time. Secondly, constantly changing nutrients. Third, prevent shaking and knocking. Fourthly, what we can now teach culture is very limited things. We do not know how far we can go . ”
Biomedical research in Russia is poorly developed, and the market for specialists is small. But, according to Michael, a person who has worked in this field can easily find application in other countries. It can go, for example, towards pharmacology, that is, to develop neurohybrid systems (testing of various drugs; research of drugs related to the treatment of degenerative diseases).
On the other hand, there is neuroscience research. Or applications related to machine learning. In any case, all of these areas, in addition to basic research, have a large field for applied research.
You can see Michael’s laboratory and listen to it here.
To get to the institute's territory, you need to try hard at any laboratory: negotiations with the press service, an exact list of equipment (don't let Gd have an extra laptop) and a mandatory dress code - white coats, outstanding at the entrance. But it was worth it, and this is why: in general, PostScience has a long relationship with Mikhail Burtsev, he often comes to the editor to write down new videos, but it was great to communicate with him in his natural habitat - all scientists are right in the laboratories and show simple at first glance, but amazing things.
')
Burtsev Laboratory is engaged in the construction of neurointelligence systems. That is, trying to solve the complex problem of building artificial human intelligence, breaking it up into stages. First you need to understand how simpler networks work. Then, having understood the principles of how cells interact with each other, scale them at the level of the animal's intelligence.
Mikhail Burtsev: “It is important that we try to combine biological experiments and theoretical work in the field of neurointelligence systems in one laboratory. That is, we build computer models. There is some agent living in an intellectual environment. In order for this agent to solve some tasks in the environment that surrounds him, he must perform certain sequences of actions, receiving a reward for this. The agent needs a neural network that will control his actions.
The neural network will see what has happened in the environment. She must learn to solve problems in this environment. We have theoretical assumptions about how this should work. On the other hand, in order to test theoretical positions and, possibly, to extract some principles from real systems, we want to investigate what is happening in real networks in living neurons when they are being trained. Today came an important moment in the field of experimental neurobiology. A critical mass of technologies has been accumulated, which makes it possible to investigate neural networks alive in great detail. ”
The construction of neuro-intellectual systems as a direction in science has arisen within the research on the creation of artificial intelligence. It is due to the fact that in research on artificial intelligence there are two alternative approaches to how to create this artificial intelligence. One approach comes from psychology and our understanding of how a person thinks and thinks. In the laboratory, Burtsev is trying to reproduce in a computer program those intellectual abilities of a person that can be learned from psychology. The second approach comes from below. Since it is clear that our intellect is based on the interaction and collaboration of many billions of cells in the brain, we can try to model these nerve cells and build artificial neural networks of them that will perform intellectual tasks.
Both approaches have their own advantages and disadvantages. For example, the first approach comes from cognitive sciences. It allows you to quickly make applications for tasks that are associated with human speech activity, with the analysis of texts, to make systems that will communicate with a person in a natural language. But there is a problem in it: human consciousness is the most complex object of all that exists, and building a normal model of such an object is not easy, because it is difficult to understand which principles are necessary. And to make a complete normal model that would have at least the child’s intelligence is almost impossible now.
On the other hand, if you simply assemble a human brain from neurons, it will not work either and will not start thinking like a human. Knowledge of neuroscience today is not enough to build such a complex system.
What methods of brain research exist?
You can put electrodes on your head, but pulling out information about how the brain works is very difficult. As studies of the last decades show, it is the intellectual component (brain training and memory) that is associated with the work of individual specific cells. When electrical activity is removed from the surface of the skull, a huge number of cells are summed up, so it is impossible to understand how the specific activation of cells in the brain has changed in specific tasks.
Another way to help see real cells in the brain is to implant an electrode in the head. So you can register individual cells.
MB: “There were a lot of experiments when it was shown that neurons can be found that are active only when a person sees certain photos. Or when he remembers something. Electrodes were implanted to the person in order to conduct further surgical intervention to stop epileptic seizures. For this you need to investigate in which place you need to cut a piece of the brain. Having such electrodes, we can check on a person how his brain works. My favorite video is when a person is shown cutting different segments of video clips. There, and the NBA, the clip of Madonna, a cartoon about the Simpsons, a picture of how the car enters Hollywood. We see in the video that the person is watching the video, and the Simpsons are shown to him, he begins to activate one cell. We hear knocking, cell discharges. In the second part of the video, a person is asked to tell what he saw. The man begins to say: "I saw the sign of Hollywood" and something else. Then he pauses. We hear this cell begin to activate. He laughs and says: "I also saw the Simpsons." This means that this neuron works both at the moment of the perception of this video and when it remembers it. He loses in his head how he sees this video. It is clear that we can learn a lot from these experiments. On the other hand, in order to understand how this all works at the level of the organism and of purposeful behavior, much more needs to be done. Imagine there is a composer who composes a symphony. The mystery is how a change in the behavior of individual cells in the brain leads to the fact that he comes up with a symphony that will appeal to hundreds of thousands of people? That is, how the change of thousands of cells at the same time led to the fact that he wrote some great work? How do these cells even know what this composer wants? What do these cells see?
These cells see some chemical signals that other cells send them. That's all. How to decode these signals? Do they come together to create this symphony? Or how to invent something or prove some theorem? This is the mystery over which we are fighting, trying to understand how this is happening. ”
To do this, you need to understand how several cells work at once simultaneously. Ideally, you need to have a brain in which you can see all the cells and then you can turn to neuronal cultures.
MB: “Neuronal culture is the cells that we took from the brain and put in a cup (in a nutrient medium). They grow there. And we know that they are starting to form networks there and show spontaneous electrical activity. We use the following cups in our experimental work. This is a round chamber in which we plant cells. There is also a nutrient medium. And along the edges of the square matrix there are contacts, to which electrodes are located, located in the center of the cup. The bottom of the cup is covered with a special solution that makes the cells adhere to the bottom. Thus, cells do not float in the nutrient medium. If this coverage is not present, the cells will begin to form clusters, because they have such a genetic need. They stick together and form the so-called neurospheres that float there. And to avoid this, we cover the bottom. And the cells stick to the bottom. And if they float, we cannot measure their electrical activity, because we cannot lead the electrode.
And so the cells are spread by a monolayer, or in two - three cells. And then all these cells are visible under the microscope. We will be able to land 5 -10 thousand cells there. Although, of course, in such a number of cells is difficult to understand. But still, comparing with the millions of cells in the brain, it is much easier. Accordingly, we can register what happens to these cells in terms of their electrical activity. This is important because the cells come together due to electrical activity. With the help of electricity, they emit chemical messages to each other. And when the cell is activated, we know that one cell passed something to another.
Obviously, if we want to build an intellectual system and understand how it solves some problem, then we will not achieve anything by simply placing the cells in a cup. There is some kind of activity, and no more. If you look in a microscope at the cells planted in a cup, they all lie first in separate circles. These are dissociated cultures.
A dissociated culture, when we take cells out of the brain, sever ties between them and plant them to the surface. They first grow on the surface. And after two or three days, they release processes and begin to connect with each other, feel each other and make contacts. When they have already established these contacts, they send messages to each other through these contacts. We can say that they have a fear of being alone. And they try to talk to each other. It turns out a social network.
The problem is that we do not understand what is happening there. United cells live some kind of life. We do not know what they think about when they are united in the brain. But we want to understand how neural networks adapt to something. To understand this, we need to set a task for the neural network that it will solve. Then we will be the result of solving this problem. And we will understand that in this moment the network has been trained, but not yet in this. ”
Who are all these people
Burtsev, like many others, has a disciplinary area, and it also combines biology, mathematics and physics. More precisely, even people from Computer Science who can think abstractly, biologists who can do something with their hands. Often people move from one area to another.
MB: “For example, if I do machine learning, I know what an artificial neural network is. I can program them, build a program that will recognize images. But, if I want to develop these algorithms, to see where they can move on, I need to know how real neurons work. In general, feel what it is.
The experience that I get when working with real neuronal cultures and with real neural networks can be invaluable. In my experience, there are more specialists from the theoretical side in the study of neuro-intellectual systems among physicists and specialists from Computer Science. It would seem that if we have an interdisciplinary audience, then there should be an opposition between biologists, on the one hand, and physicists and mathematicians, on the other. In fact, it does not exist as such. Those biologists who are engaged in experiments need methods for analyzing and processing data. But they may lack skills from an instrumental point of view. Here they turn for help to physicists and mathematicians who understand how this data can be better processed and presented.
That is, physicists and mathematicians in this case do not provide conceptual, but technical assistance. And vice versa. When a physicist tries to do an experiment with cells, it is obvious that he needs help, telling what to do and helping to do this experiment. That is, there is mutually beneficial cooperation. ”
And then where
The main application of the study of neurointelligence systems are pharmacology and medicine. The second is an attempt to build intelligent systems. The last direction is the activity of the laboratory Burtsev.
MB: “The first layer of this activity is to teach the cells to control a real robot. We will launch a robot under water, collect volcano or debris from the street, solve the migration problem. But it is not yet clear how to implement this from the point of view of technology. First, we need to maintain sterile conditions all the time. Secondly, constantly changing nutrients. Third, prevent shaking and knocking. Fourthly, what we can now teach culture is very limited things. We do not know how far we can go . ”
Biomedical research in Russia is poorly developed, and the market for specialists is small. But, according to Michael, a person who has worked in this field can easily find application in other countries. It can go, for example, towards pharmacology, that is, to develop neurohybrid systems (testing of various drugs; research of drugs related to the treatment of degenerative diseases).
On the other hand, there is neuroscience research. Or applications related to machine learning. In any case, all of these areas, in addition to basic research, have a large field for applied research.
You can see Michael’s laboratory and listen to it here.
Source: https://habr.com/ru/post/202952/
All Articles