From the living brain to artificial intelligence
Over the past five years, a large wave of publications, speeches, fears, as well as hype - discussions, hype and aggressive advertising has risen around artificial intelligence. This indicates primarily speculation and profanity. News portals vying with each other tell how the neural network learned another amazing trick. It should be understood that an artificial neural network is not a model of the brain. Neurophysiologists note that the mechanism of the brain is still poorly understood, and its model is far from mathematical formalization.
In 2013, the European Commission evaluated The Human Brain Project, a research project, and allocated a $ 1 billion grant for research. The goal of this project is to create a fully functional computer model of the human brain. Present the complexity of the brain: how neurons connect and interact with each other, how memories are formed, and decisions are made. In 2019, at the European Conference, a presentation was made by the neurobiologist, director and founder of The Human Brain Project, Henry Markram. He and the team showed images of the brain that were illuminated on the tomograph from different sides, emphasizing that the more they went deeper into the brain, the less it became clear how it works. All this time they moved along the path of miniaturization, but, obviously, we need a macro-model of the brain. Here we can draw an analogy with the well-known metaphor: no matter how much the ant crawls inside the TV, it still does not understand what kind of device it is.
Chatbots speak but don’t understand
The first conversational program "Eliza" was developed back in the 60s. Her work was like a psychotherapy session. An example dialogue might look like this:
- Hello. How are you feeling?
- My head hurts.
- How often does this happen?
“Every time I talk to my mother.”
“You have a problem with your mom.” Tell us more about this.
The program contains a set of blanks. Identifies the keyword in each sentence and creates the illusion of conversation, without any meaning at all. Thus, Eliza could skillfully maintain a conversation for several hours, making a good impression on people.
Today's Alice from Yandex is the same Elisa, with a connected service platform and much richer vocabulary. All the intelligence of such chat bots comes down to retraining. If we mark out for them the answers are bad, unsuitable, then it will weed them out in the future, and leave good ones that are suitable. The real reasonableness of the program, its ability not only to determine what is shown in the picture or to convert oral speech into text, but to understand the meaning of information and think, will only appear when connected to a knowledge base. This is a certain understanding of the world order. No program can do this yet. Artificial intelligence working with a knowledge base is called Symbolic AI (symbolic AI).
Now two directions are actively used: artificial neural network and machine learning. Their development is not carried out in scientific laboratories under the guidance of scientists, but in IT companies. Programmers and developers of artificial neural networks admit that there is nothing thinking except the brain. Therefore, if they want to create an intelligent system, then they are forced to simulate the brain, because there is nothing more. But a cardinal breakthrough will be for those who are able to cross models of knowledge representation and machine learning.
By the way, some companies are already actively using robots for job placement. They are especially good at mass screenings, which receives more than 1000 resumes. The robot analyzes the questionnaires and evaluates the candidates for relevance, but the final decision is still made by the person. Scientists at the Georgia Institute of Technology have found out in which case people willingly trust the robot. This happens if they are informed in advance that it is designed to perform a specific task. Such robots also conduct weekend interviews. People have a more frank dialogue with the machine about the reasons for dismissal and about what difficulties they encountered in the process of work. Thus, robots collect very good, and most importantly significant, feedback. In addition, people are afraid of lying to a robot. They think cars have more information. For example, the two main questions for selecting storekeepers were as follows: do you have a criminal record and do you drink alcohol? And people without hiding answered these questions to the robot.
What is wrong with drones
Modern unmanned vehicles are imperfect and have several disadvantages. The most obvious lies in machine vision. The issue is not even in the technical sphere; the problem of machine vision has not been solved conceptually. Cameras and sensors process certain signals with great speed, as a result of which binary patterns are formed in the computer's memory, which must be interpreted. This is the main difficulty. If, for example, a pedestrian crosses the road, then any driver easily considers this. Not because his vision is better than that of a drone’s camera, but because the driver has a model of a moving person in his mind. Vision models do not yet exist. It is not known how a person processes visual information. Modeling devices are being designed that allow with a certain degree of confidence to recognize a number of objects.
Imagine that you asked a person who does not understand anything in cars to bring the wishbone. Obviously, he will not be able to do this. Because he is not a car mechanic, and he does not have a lever model in his head. But if you say that this is a piece of iron of a curved shape - a person will bring it. Everyone knows what a piece of iron is, and understands what a “curved shape” looks like. Even after explanation, the robot will not cope. To do this, you need to train him, laying down the concepts of all these objects from different angles and with different degrees of access.
While in the office, a person does not have to move the chair away from the table to identify the item. Just look at the back. Only because people have a complete model of the chair, and we, in one part of it, recreate the whole. We know that a chair is an integral attribute of an office interior. Therefore, from a huge selection of objects that can be recreated from this fragment, we select one. There are no universal programs that could do this. No artificial intelligence now contains a model of the world and knowledge that allows it to unambiguously interpret the picture that it sees.
Suppose, somewhere behind the bushes a man hid, ready to run out onto the road. Part of his body is hidden behind dense foliage. The visible part will not be enough for the machine to be able to consider this as a potential danger. Moreover, if you see an elephant hiding behind the bushes, you probably decide that this is a man in the mask of an elephant, because in our area elephants are not found. Well, the car certainly does not have this knowledge.
The problem of interpretations is approaching its resolution, in any case there is progress. In the machine learning algorithm on huge samples, you can put 100 cars, and it is clear that 101, even of a different brand, with a high probability, the program will recognize correctly. Although this car will not be specifically laid down in it. It is also worth noting that for the training of the program it is important to observe the variety of selection. If, for example, you only learn the program in sedans, and then a convertible drives up, then it probably will not recognize it, because there were no cars without a roof in the sample.
The second challenge is an ethical dilemma in which a machine has to make a moral choice. Let's say unmanned vehicles carry a passenger. A man jumped out to meet, and the only way to avoid a collision was to drive into a nearby pillar. The car must decide who to sacrifice - those who are inside the car, or those who rushed under the wheels. This is an absolutely insoluble task for her. Already there are first accidents and even victims. In Arizona, an unmanned Uber-owned SUV hit a pedestrian to death. Their fault is that they released a crude algorithm that did not pass all the tests on the track.
The problem of drones is, first of all, the lack of basic ideas about the world. People always make decisions in context. No artificial intelligence has a complete picture of the world that a person has at the age of 18. For this reason, driver's licenses are issued precisely from the age of 18 [in Russia], although the vision is well formed already at 14. Before reaching adulthood, a person cannot make informed decisions, including ethical and emotional ones.
We are dealing with very young and immature algorithms that require refinement. They are able to work properly, but exclusively under human control.
Polygraphic Artificial Intelligence
In the artificial intelligence market, the world's leading companies are actively investing in the field of emotional computing. This technology reveals the most invisible changes in facial expressions. The program selects a certain number of points on the face and compares it with the database of photographs in which the emotion is already recognized.
Microsoft said that they have a similar algorithm, but it will not give it to the government. This is a very dangerous tool that can be used against humans. Imagine that you enter the office of the boss, and he sees that you hate him and think badly of him. There is a serious ethical problem.
There will be many more similar recognizer programs. Beautiful, wonderful including. They will be useful, especially in medicine. Cars are already helping doctors diagnose diseases. Let's say the American IBM Watson program makes diagnoses by analysis better than some novice doctors. As a training sample, six thousand case histories were laid in it. This is a titanic work and, of course, a lot of money.
Does the machine know how to compose
The machine is not able to generate something qualitatively new. Original lyrics, verses, musical compositions - all this is based on the principle of permutation, or, more simply, permutation. As for poetry, the algorithm of actions is as follows: the program finds common moments and certain combinations. He takes keywords, adds the words of other authors to them and twists them to the desired rhythm. The Russian language is complex, but if you do not care about rhyme, then you can “compose” a white verse.
With music it's still easier. The program determines which chords are most characteristic of a particular artist and uses them, but in a different sequence. At the heart of this "creativity" lies the shuffling of the scale and the rejection of overt cacophony.
There were programs that composed folk tales. To do this, we conducted an analysis of the scenarios using the book by Vladimir Propp “Morphology of a Tale”. It turned out that the model of the development of events includes several obvious elements: fairy tales always have positive and negative characters. There is also a donor, a road, an obstacle, and a happy ending at the end. A distinctive feature of such tales was eclecticism. The program combined the heroes and their actions, telling, for example, about Ivan the Fool, who is fighting with the Serpent Gorynych. An absolute minus was the poor language of the story. Listeners got bored. A living storyteller will always add humor, and somewhere amusing speech speed. For the same reason, the program cannot translate metaphors or phraseological units into other languages. Online translators work with large volumes of information. It is worth noting that the quality of their translation in recent years has improved markedly. Now they do not translate each word individually, as they did before, but look for the whole sentence somewhere in the texts. The biggest difficulty arises with semantics. This requires a knowledge base to understand the meaning of expressions. For example: chickens are ready for dinner. It is unclear how to translate this phrase correctly if you do not know where the person is at that moment - at the chicken farm (Chickens are ready for dinner) or in a restaurant (Chickens are prepared for lunch).
Without human knowledge, no intellectual activity is possible. Knowledge cannot be replaced by anything. The question is how to digitize these skills. How to turn experience into some meaningful scheme that could be shared with the machine.