Artificial intelligence cybernetics and neural networks. Neural networks. Promising areas for the implementation of neural networks

Another area of research in artificial intelligence is neural networks. They were designed to resemble the natural neural networks of the human nervous system.

Artificial neural networks

The inventor of the first neural computer, Dr. Robert Hecht-Nielsen, defined a neural network as follows: “A neural network is a computing system consisting of a number of simple, highly interconnected processing elements that process information by their dynamic response to external influences.”

Basic structure of artificial neural networks (ANN)

The idea of an ANN is based on the belief that it is possible to imitate the functioning of the human brain by creating the necessary connections using silicon and wires such as in living neurons and dendrites.

The human brain is made up of 100 billion nerve cells called neurons. They are connected to other thousands of cells by axons. Stimuli from the external environment or signals from the senses are received by dendrites. These input signals create electrical impulses that quickly travel through the neural network. The neuron can then send messages to other neurons, which may send the message further or may not send it at all.

Artificial neural networks consist of several nodes that mimic the biological neurons of the human brain. Neurons are connected and interact with each other. Nodes can accept input data and perform simple operations on the data. As a result of these operations, data is transferred to other neurons. The output for each node is called its activation.

Each link is associated with a weight. ANNs are capable of learning, which is carried out by changing the weight value. The following figure shows a simple ANN:

Types of artificial neural networks

There are two types of artificial neural network topologies - feedforward and feedback.

The flow of information is unidirectional. The unit transmits information to other units from which it does not receive any information. No loop feedback. They have fixed inputs and outputs.

Here, feedback loops are allowed.

How artificial neural networks work

The topology shows circuits, each arrow representing a connection between two neurons and indicating a path for the flow of information. Each connection has a weight, an integer that controls the signal between two neurons.

If the network produces a “good” and “needed” output, then there is no need to adjust the weights. However, if the network produces a “bad” or “unwanted” output or error, then the system adjusts its weights to improve subsequent results.

Machine learning in artificial neural networks

ANNs are capable of learning, and they must be trained. There are several teaching strategies

Training - involves a teacher who submits a training sample to the network to which the teacher knows the answers. The network compares its results with the teacher's responses and adjusts its weights.

Unsupervised learning is necessary when there is no training set with known answers. For example, in clustering problems, i.e. dividing a set of elements into groups according to some criteria.

Reinforcement learning is a strategy built on observation. The network makes decisions by observing its surroundings. If the observation is negative, the network adjusts its weights to be able to make the different decisions needed.

Backpropagation algorithm

Bayesian networks (BN)

These are graphical structures for representing probabilistic relationships between a set of random variables.

In these networks, each node represents a random variable with specific offers. For example, in a medical diagnosis, a Cancer node represents the proposition that a patient has cancer.

The edges connecting the nodes represent probabilistic dependencies between these random variables. If of two nodes, one influences the other node, then they must be directly connected. The strength of the relationship between variables is quantified by the probability that is associated with each node.

The only limitation on arcs in BN is that you cannot go back to a node simply by following the direction of the arc. Hence the BNN is called an acyclic graph.

The BN structure is ideal for combining knowledge and observed data. BN can be used to learn cause-and-effect relationships and understand various problems and predict the future, even in the absence of data.

Where are neural networks used?

They are able to perform tasks that are simple for humans but difficult for machines:

Aerospace - aircraft autopilot;

Automotive - automotive guidance systems;

Military - target tracking, autopilot, signal/image recognition;

Electronics - forecasting, fault analysis, machine vision, voice synthesis;

Financial - real estate valuation, credit consultants, mortgage, trading company portfolio, etc.

Signal Processing - Neural networks can be trained to process the audio signal.

Japanese algorithm wrote a book“The Day the Computer Wrote a Novel.” Despite the fact that people helped the inexperienced writer with the characters and storylines, the computer did a great job - as a result, one of his works passed the qualifying stage of a prestigious literary award. Neural networks also wrote sequels to Harry Potter and Game of Thrones.

In 2015, the AlphaGo neural network, developed Google team DeepMind became the first program to beat a professional Go player. And in May of this year the program beat the strongest Go player in the world, Ke Ze. This was a breakthrough because for a long time It was believed that computers did not have the intuition necessary to play Go.

Safety

A team of developers from the University of Technology Sydney has introduced drones for patrolling beaches. The main task of drones will be searching for sharks in coastal waters and warning people on beaches. The analysis of video data is carried out by neural networks, which significantly affected the results: the developers claim that the probability of detecting and identifying sharks is up to 90%, while an operator viewing video from drones successfully recognizes sharks only in 20-30% of cases.

Australia ranks second in the world after the United States in the number of shark attacks on people. In 2016, 26 cases of shark attacks were recorded in this country, two of which resulted in death.

In 2014, Kaspersky Lab reported that their antivirus registered 325 thousand new infected files every day. At the same time, a study by Deep Instinct showed that new versions of viruses are practically no different from previous ones - the change ranges from 2% to 10%. The self-learning model developed by Deep Instinct, based on this information, is capable of high accuracy identify infected files.

Neural networks can also look for certain patterns in how information is stored in cloud services, and report detected anomalies that could lead to security breaches.

Bonus: neural networks guarding our lawn

In 2016, 65-year-old NVIDIA engineer Robert Bond was faced with a problem: his neighbor's cats were regularly visiting his property and leaving traces of their presence, which annoyed his wife, who was working in the garden. Bond immediately rejected the too unfriendly idea of building traps for uninvited guests. Instead, he decided to write an algorithm that would automatically turn on garden sprinklers when cats approached.

Robert was faced with the task of identifying cats in a video stream coming from an external camera. To do this, he used a system based on the popular Caffe neural network. Every time the camera observed a change in the situation on the site, it took seven pictures and transmitted them to the neural network. After this, the neural network had to determine whether a cat was present in the frame, and, if the answer was yes, turn on the sprinklers.

Bond yard camera image

Before starting work, the neural network was trained: Bond “fed” it 300 different photos cats. By analyzing these photographs, the neural network learned to recognize animals. But this was not enough: she correctly identified cats only 30% of the time and mistook Bond's shadow for a cat, as a result of which he himself ended up wet.

The neural network worked better after additional training on more photographs. However, Bond warns that it is possible to train a neural network too much, in which case it will develop an unrealistic stereotype - for example, if all the images used for training are taken from the same angle, then artificial intelligence may not recognize the same cat from a different angle. Therefore, competent selection of the training data series is extremely important.

After some time, the cats, who had learned not from photographs, but from their own skin, stopped visiting Bond's site.

Conclusion

Neural networks, a technology from the middle of the last century, are now changing the way entire industries operate. The reaction of society is ambiguous: some are delighted with the capabilities of neural networks, while others are forced to doubt their usefulness as specialists.

However, not everywhere it comes machine learning, it displaces people. If a neural network makes diagnoses better than a living doctor, this does not mean that in the future we will be treated exclusively by robots. Most likely, the doctor will work together with the neural network. Similarly, the IBM supercomputer Deep Blue beat Garry Kasparov at chess back in 1997, but people from chess have not disappeared anywhere, and famous grandmasters still appear on the covers of glossy magazines.

Cooperation with machines will bring much more benefits than confrontation. Therefore, we have compiled a list of materials in open access that will help you continue your acquaintance with neural networks:

Almost always, when the term “AI” appears in the popular press, it is just a fancy term for neural networks to make the article clearer for the reader (and also to lure more readers). Apparently, such a culture has already developed among observers: if you see a neural network, write that it is “AI”.

In terminology: AI is a program that can learn (gain knowledge and experience independently in the process of its work) and effectively use its experience in the future to better perform a certain task (the task for which this program was created). If this specialized program(for example, for playing chess) - she is called “weak” AI because her ability to understand (gain experience) is specially created and adapted for chess. Many weak AIs have already been created. For example, Alpha Zero, based on neural networks, and taught itself to play chess from scratch, is now a contender for the title of the world's strongest chess player (it plays at approximately the same level and perhaps even surpasses the best chess programs created by classical programming based on chess theory). The term "strong AI" (or "AI general purpose") is reserved for a hypothetical program that is capable of independently learning various tasks (that is, it does not have special programming for a specific task). At the moment, there is no such program and is not expected to appear soon (in the next 5 years). Developments in this direction are being carried out (including by the authors of Alpha Zero mentioned here).

Note that in the term AI there is no mention of neural networks. Because this term does not describe any technology, not a tool, not a means. The term AI describes the end result (the ability to learn and use what has been learned). That is, AI can be created on the basis of neural networks, or perhaps without them (although, most likely, these will indeed be neural networks).

On the other hand, the term “neural network” describes precisely the technology (idea, approach) of programming. The idea is that instead of programming all the actions of the program command by command (as in classical programming), create some basic structure with the most general ideas about what it will have to work with. Moreover, this structure is based on a huge number of parameter numbers (millions (billions? trillions?)), but they are deliberately left blank (initially there are some approximate average values and a little garbage). In many ways, this structure is similar in operating principle to the work of the human brain (which is why it is called " neural network"). Then, through a huge number (from hundreds of thousands to millions), so to speak, “practical tasks” (the program performs the work and then the success of this program is assessed), step by step it is determined which specific values of all these parameters will lead to the best result. This process of searching for the best parameters is learning, gaining experience. Since the very idea of a neural network is that it must learn a lot, any working neural network is to some extent an AI (if it is at least some useful task decides).

So the term AI describes an idea. AI can be based on the principles of neural networks, or maybe on some others. The term "neural network" describes the technology (which so far shows to be the most promising for the future of strong AI).

Less than a year ago I read an interview with the head of Deep Mind (considered the leading (or at least one of the leading) company in this field). He mentioned that human-level AI, in his opinion, will be created “in tens of years” (quote). It is impossible to say more precisely now.

Let me remind you again that at the moment, even a “simple” (although this is, in fact, not simple) general-purpose AI that would have self-awareness and any interesting analytical abilities has not been created. Therefore, it is too early to talk about this now.

Although in general it seems that in general terms it is already clear how this can be done, which way to go. Simple neural networks produce simpler results. Complex multi-layer neural networks, and also recurrent ones (when the outputs of this network are re-directed to the inputs, allowing the network to first make the first conclusion, and then think about this conclusion, and then think about the decisions that appeared during thinking, and so on and so forth, gradually moving from momentary details to the big picture, each time abstracting more and more) give much more impressive results. Observing their work, the learning process, one gets the impression that they behave (make decisions) almost exactly the same as a person who is learning the same thing (for example, learning to play chess). Neural networks come up with approximately the same ideas in approximately the same sequences as a human. There is an evolution of ideas from primitive momentary (even “right now”) to complex, intricately intertwined ideas that take into account many nuances.

For example, some professional players in chess, watching Alpha Zero play, they say that its play is not like that of a typical chess program. One of the professionals compared her game with the game of some very intelligent alien creature (Garry Kasparov, one of the best players in the history of chess, for example, said that her game is something between the game of a good computer program and a good human chess player).

So it seems that the direction has been chosen correctly: recurrent neural networks. To achieve humanoid AI, it is likely that several new types of networks will need to be developed at once and successfully combined together. This requires a huge amount of calculations, experimentation, trial and error. And also a huge amount of computing power. The process of training such a colossal neural network will require many computers. So many. Even training a relatively simple network like Alpha Zero required resources equivalent to dozens of years of work personal computer. A neural network capable of thinking like a person will require hundreds (hardly so little), thousands or even millions of times more calculations for its training. However, such computing power is not an insurmountable barrier. The main thing is that there is energy, there is enough of it, you can add as many processors and memories as you like, there is no problem here.

Answer

Comment

They were designed to resemble the natural neural networks of the human nervous system.

Artificial neural networks

Basic structure of artificial neural networks (ANN)

Each link is associated with a weight. ANNs are capable of learning, which is carried out by changing the weight value. The following figure shows a simple ANN:

Types of artificial neural networks

There are two types of artificial neural network topologies - feedforward and feedback.

The flow of information is unidirectional. The unit transmits information to other units from which it does not receive any information. There is no feedback loop. They have fixed inputs and outputs.

Here, feedback loops are allowed.

How artificial neural networks work

Machine learning in artificial neural networks

ANNs are capable of learning, and they must be trained. There are several teaching strategies

Unsupervised learning is necessary when there is no training set with known answers. For example, in clustering problems, i.e. dividing a set of elements into groups according to some criteria.

Backpropagation algorithm

Bayesian networks (BN)

These are graphical structures for representing probabilistic relationships between a set of random variables.

In these networks, each node represents a random variable with specific offers. For example, in a medical diagnosis, a Cancer node represents the proposition that a patient has cancer.

The only limitation on arcs in BN is that you cannot go back to a node simply by following the direction of the arc. Hence the BNN is called an acyclic graph.

Where are neural networks used?

They are able to perform tasks that are simple for humans but difficult for machines:

Aerospace - aircraft autopilot;

Automotive - automotive guidance systems;

Military - target tracking, autopilot, signal/image recognition;

Electronics - forecasting, fault analysis, machine vision, voice synthesis;

Financial - real estate valuation, credit consultants, mortgage, trading company portfolio, etc.

Signal Processing - Neural networks can be trained to process the audio signal.

The key to the success of artificial intelligence development is neurobiology. How exactly scientists are trying to reproduce the work of the human brain and what is special about the work of neural networks is in the material of “Futurist”.

Neuroscience and artificial intelligence

“The future of artificial intelligence is in neurobiology,” says the founder of Google DeepMind, doctor of neurobiology Demis Hassabis (Demis Hassabis) in an article published in the journal Neuron. Hasabis launched his London-based company DeepMind to create technical analogue of human intelligence, and Google bought his company for more than $500 million in 2014. Last year, AlphaGo, a program developed by DeepMind, beat the world champions in the logic game Go. In collaboration with OpenAI, a non-profit AI research institute supported by Elon Musk , the company is also working on creating machines with more advanced intelligence capabilities.

All of DeepMind's artificial intelligence algorithms are based on concepts first discovered in our own brains. Deep learning and reinforcement learning - two pillars of modern AI - are the result of a free translation of the model of the biological neural connections of the human brain into the language of formal mathematics. Deep learning is actually just a new name for a more than 70-year-old approach to artificial intelligence known as neural networks. Neural networks were first proposed back in 1944 Warren McCullough (Warren McCullough) and Walter Peets (Walter Pitts), two University of Chicago researchers who founded what is sometimes called the first department of cognitive science in 1952.

Neural networks were a major area of research in both neuroscience and computer science until 1969, but interest in them has since waned. The technology began to make a comeback in the 1980s, but went into eclipse again in the first decade of the new century and returned almost immediately in the second, thanks largely to the increased processing power of graphics chips.

Neural network device diagram

Features of the operation of neural networks

Neural networks are a machine learning tool in which a computer learns to perform a specific task by analyzing learning examples. Typically, these examples are pre-labeled by hand. For example, an object recognition system could store thousands of labeled images of cars, houses, cups, etc., and be able to find visual patterns and features of these images to later associate them with specific labels. Simply put, this is also how children learn - for example, a child is shown different red objects, so that in the future he can independently associate this “label” with all red objects.

However, in order to develop even a remote technical analogue of the connections of our brain, the creation of a complex mechanism is required. Neural networks are made up of thousands or millions of simple but densely interconnected information processing nodes, usually organized in layers. Various types networks vary depending on their number of layers, the number of connections between nodes, and the number of nodes in each layer. Most modern neural networks are organized into layers of nodes in which data moves in only one direction. A single node can be connected to multiple nodes in the layer below it from which it receives data, and to multiple nodes in the layer above it to which it sends data.

Network training example

A node assigns a number to each of its incoming connections, known as a “weight.” When the network is active, the node receives another data item, another number from them and multiplies it by the already given weight, and then adds the values received from all the inputs together, getting one number. If the number exceeds a threshold, the node "fires", which in modern neural networks usually means sending the number - the sum of the weighted inputs - across all of its outgoing connections.

In training mode, all weights and thresholds of the neural network are initially set to random values. The training data is fed to the bottom layer - the input layer - and passes through subsequent layers, multiplied and added, until it reaches the output layer. During training, the weights and thresholds are continually adjusted until training data with the same labels achieve similar results.

The future is already here

The impressive results of improving the performance of neural networks and spreading the use of technology are not limited to the victory of AlphaGo and laboratory AI research. If you still associate the phrase “self-learning machines” with the world of science fiction and horror films about the uprising of robots, then welcome to the future.

IN last years the most effective artificial intelligence systems - in areas such as autonomous driving, speech recognition, computer vision And automatic translation- were developed thanks to neural networks. We may not even notice it ourselves, but self-learning neural networks are already ingrained in our daily lives. So, for example, the translation suggested to you in your Facebook feed is no longer done automatically by searching for every single word in dictionary. Now the company has launched a neural network that translates entire sentences, producing increasingly literate, coherent text. Already, the accuracy of translations on social networks has increased by 11%.

Neuron model processed in Prisma application

A separate wave of interest in the technology of ordinary people in Russia was caused by the appearance of the Prizma application, which turns ordinary photographs into likenesses of famous works of art. It doesn’t matter whether you have used this application or, on the contrary, were perplexed about its abuse by social network users - it is worth noting the creativity of its creators. The peculiarity of this seemingly ordinary photo processing tool was precisely that the program worked on the basis of neural networks, using the patterns of various painting styles to create new “masterpieces.”

However, even the simplest neural networks take up a lot of memory and consume huge amounts of energy, so they usually run on servers in the cloud, where they receive data from desktop or mobile devices, and then send back the analysis results.

To address this problem, last year MIT associate professor of electrical engineering and computer science Vivienne Sze and her colleagues unveiled a new energy-efficient computer chip optimized for neural networks that could enable powerful artificial intelligence systems to run locally on mobile devices. .

In addition, they developed an analytical method that can determine how much energy a neural network consumes when running on a certain type of hardware. They then used the technology to evaluate new methods for bypassing neural networks so they could work more efficiently on handheld devices.

However, Hassabis argues that this is not enough. The goal that researchers now set for themselves is to create a universal AI with the ability to think, reason and learn quickly and flexibly, artificial intelligence capable of understanding the real world and imagining a better one.

To achieve this, it is necessary to study the workings of the human mind more closely, since it is the only evidence that such an intelligent system is possible in principle.

AI training problem

Depending on their specific tasks, machine learning algorithms are tuned using specific mathematical structures. Through a million examples, artificial neural networks learn to fine-tune their connections until they reach an ideal state that allows them to perform a task with the highest possible accuracy.

Because each algorithm is completely tailored to a specific task, retraining for a new task often erases previously established connections. Thus, when the AI learns a new task, it completely overwrites the previous one.

The dilemma of continuous learning is just one problem in artificial intelligence. Others haven't even been so precisely defined yet, but perhaps they will prove more significant in creating flexible, inventive minds like ours.

For example, the problem of embodied cognition - as Hassabis explains, is the ability to create knowledge and abstract thoughts based on independent sensory interaction with the world. It's a kind of common sense that people have, an intuition that is difficult to describe, but which is extremely useful for solving the everyday problems we face.

Traits such as imagination are even more difficult to program. This is where AI, limited to one specific task, is really bad, Hassabis says. Imagination and innovation are based on the models we have already created about our world - and imagining new scenarios from them. These are very powerful planning tools, but their research for AI is still in its early stages.

Scientists note that when solving problems of neural networks, they turn to neurobiology and physiology of living beings. Thus, recent discoveries show that the hippocampus - part of the brain's limbic system, which is responsible for memory - “replays” our experiences and memories in fast forward during rest and sleep. This allows the brain to "relearn from successes and failures that have already occurred in the past," Hassabis says.

AI researchers took up this idea and implemented a rudimentary version of the algorithm - resulting in powerful, experience-learning neural networks. They compare current situations with previous events stored in memory and take actions that previously resulted in success or reward.

But the best is yet to come

The advent of brain imaging and genetic bioengineering tools offers unprecedented insight into how biological neural networks organize and connect to solve problems. As neuroscientists work to solve the “neural code”—the basic computations that support brain function—AI researchers have an increasingly broader set of tools to study.

It’s worth noting that it’s not just AI that has something to learn from neuroscientists—the benefits are mutual. Modern neuroscience, in all its powerful means imaging and optical genetics, has just begun to understand how neural networks support more high level intelligence.

“Neuroscientists often have a fairly vague understanding of the mechanisms underlying the concepts they study,” says Hassabis. Because AI research is based on rigorous mathematics, it can offer ways to clarify these vague concepts into real-world hypotheses.

Of course, it is unlikely that AI and the brain will always work in the same way. But we can think of AI as applied computational neuroscience, Hassabis says. Comparing AI algorithms to the human brain "can provide insight into some of the mind's deepest mysteries."