Neural Networks Neural Networks of Neurotechnology

Main article: Machine Learning

Neural networks are able to independently learn and develop, building their experience on perfect mistakes. By analyzing and processing information from a specific source, or from the Internet as a whole, a self-organizing system is able to create new products, not only reproducing and structuring input data, but also forming a qualitatively different result, previously inaccessible to artificial intelligence.

Neuronet (NeuroNet) is one of the supposed and most likely stages of Internet development. At the new stage of the development of the World Wide Web, the interaction of participants will be carried out on the principles of neurocompunctions, i.e. on the basis of the transfer of information about the activity of the brain. Scientists predict the formation of the Neuronet market by 2030-2040. Moreover, it is expected that at that time at least 10 Russian companies with a total capitalization of about 700 billion rubles will already function on the market.

Neural Network Security

2022: AI may be wrong even with discreet data modifications

Kryptonit specialists conducted a large-scale study of the safety of artificial neural networks ubiquitous in computer vision systems, speech recognition and deep analysis of various data, including financial and medical. The company announced this on November 29, 2022.

Experts compared the attacks on machine learning models (ML) based on artificial neural networks (INS) described in scientific articles, reproduced different implementations of attacks and talked about the 10 most visual ones.

In their study, the authors used three generally accepted scenarios in information security:

• a white-box attack involves full access to network resources and datasets: knowledge of the network architecture, knowledge of the entire set of network parameters, full access to training and test data;
• a gray-box attack is characterized by the attacker having information about the architecture of the network. Additionally, it may have limited access to data. It is attacks like a "gray box" that are most often found in practice.
• a black-box attack is characterized by a complete lack of information about a network device or a set of training data. In this case, as a rule, it is implicitly assumed that there is unlimited access to the model, that is, there is access to an unlimited number of pairs "investigated model" + "arbitrary set of input data."

Various libraries have been tested to create malicious examples. Initially, AdvBox, ART, Foolbox, DeepRobot were selected. The performance of the AdvBox turned out to be very low, and DeepRobot was very raw at the time of the study, so ART and Foolbox were in the dry residue. Experiments were carried out on various types of ML models. In its report, Kryptonite shared the most visual results obtained using one fixed model based on a convoluted neural network and five different attacks. Their implementations are taken from two libraries.

The demonstration used the MNIST database, which contains 60,000 samples of handwritten numbers, and selected the most visual malicious examples.

The number above is the absolute value of the maximum deviation of the disturbance from the original. Below the image are three numbers: maximum deviation, minimum and average. At the bottom line is the label and probability.

The study found that there is indeed a problem with the safety of neural network-based ML models. A neural network can "confidently" give an incorrect result with very small changes in the picture or other input data - so insignificant that a person is unlikely to notice them.

So, the picture on the left is the original example in which the neural network confidently recognizes the number "4." In the middle is an unsuccessful malicious example. The image is noticeably distorted, but the neural network still recognizes the four. On the right is a working malicious example. It is visually indistinguishable from the previous one, but here the threshold of perturbations has already been overcome, beyond which the neural network is lost and gives the wrong recognition result. In this case, instead of "4," it recognizes "7." In the example above, a person confidently distinguishes the number "4" in any of the three pictures, but the original images are not always quite clear.

For example, in the next picture, an undescribed zero can be visually perceived as the number "6" - the question is where to mentally continue the line. The neural network is also not sure: it shows a low probability, but correctly recognizes zero in the image. To make the INS make a mistake, you need to change only a few pixels. In this case, the value of the introduced disturbance will be of the order of 1/256, which corresponds to the value of the color resolution.

The neural network does not always manage to deceive so easily. In case of confident object recognition, you will have to generate and check many malicious examples before you can find a worker. At the same time, it can be practically useless, as it introduces too strong disturbances noticeable to the naked eye.

For illustration, Kryptonite took the most easily recognizable digit "9" from the test set and showed some of the resulting malicious examples. The illustration shows that in 8 cases out of 12 it was not possible to build malicious examples. In the remaining four cases, the researchers deceived the neural network, but these examples turned out to be too noisy. This result is related to the confidence of the model in the classification of the original example and to the parameter values ​ ​ of various methods.

In general, the experiment showed the expected results^[1]: the simpler the changes that are made to the image, the less they affect the operation of the INS. It should be emphasized that the "simplicity" of the changes made is relative: it may be a dozen pixels, but guessing which ones, and how they need to be changed, is a difficult task. There is no nail on which the CNN classification result is completely held: in general, you cannot change one pixel so that the INS is mistaken.

PGD, BIM, FGSM, CW, DeepFool methods were the most effective for the white box scenario. Regardless of implementation, they allow a successful attack with a probability of 100%, but their use implies the presence of complete information about the ML model.

Square Attack, HopSkipJump, Few-Pixel, Spatial Transformation methods assume information about the model architecture. Isolated successful examples of attacks were obtained, but the practical use of these methods is not possible. Perhaps the situation will change in the future if there are sufficiently effective implementations that stimulate the interest of researchers in these methods.

All discussed black box methods use the confidence level returned by the neural network. If you at least slightly reduce the accuracy of the returned confidence level, then (already low) the effectiveness of the methods will drop many times.

Neural networks and Bayesian models

Two popular paradigms in machine learning. The first made a real revolution in the processing of large amounts of data, giving rise to a new direction, called deep learning. The latter have traditionally been used for processing small data. The mathematical apparatus, developed in 2010, allows you to design scalable Bayesian models. This makes it possible to apply Bayesian inference mechanisms in modern neural networks. Even early attempts to build hybrid neurobayes models lead to unexpected and interesting results. For example, by using Bayesian output in neural networks, it is possible to shrink the network by about 100 times without losing the accuracy of its operation.

The neurobayes approach can potentially solve a number of open problems in deep learning: the possibility of catastrophic retraining to noise in data, self-confidence of the neural network even in erroneous predictions, non-interpretability of the decision-making process, vulnerability to hostile attacks (adversarial attacks). All these problems are recognized by the scientific community, many teams around the planet are working on their solution, but there are no ready-made answers yet.

Neural networks for creating pictures

Main article: Neural networks for creating pictures

How to use neural networks

The world has created neural networks capable of painting in any existing artistic style, confidently beating the world champion in the most difficult logical game on the planet, recording music albums and imitating human behavior in electronic correspondence. All of the above is still only a demonstration of part of the capabilities of the technology, the real application of which both in business and in everyday life, we will see in the near future.

In other words, neural networks will allow not only and not so much to replace human labor in more complex labor activities, but to become a useful tool for specialists and managers of many areas.

Vlad Shershulsky, Director of Advanced Technologies, Microsoft Rus comments: "This area finally became hot in 2016: since about 2009, there has been rapid progress in creating more and more complex, and at the same time more and more efficient, deep neural networks, and most recently we have seen impressive applications and witnessed the creation of a number of successful startups. The threshold for entering the neural network services market has significantly decreased, and projects built around the idea of ​ ​ one interesting application are implemented in a matter of months. All this gave rise to a boom in neural network startups, aroused the interest of large corporations and contributed to an increase in demand for specialists in this field, including in Russia. It is nice to note that the most important contribution to the creation of a new generation of technologies for working with natural languages ​ ​ was made by Microsoft specialists. In the famous television series Star Trek, the creation of an online translation of the spoken word was predicted in the XXII century, and we already have it today. Of course, other applications - from predicting car breakdowns and bankruptcy of counterparties to new cybersecurity tools - are also developing very successfully. "

Neural networks in radiology

Main article: Artificial Intelligence in Medicine

Neural networks in the media

2023: What neural networks can already be used in the media today

Today, more and more people understand that the future lies with neural networks, and that things can be done on them that were previously impossible. Like any innovative product - to a wide audience, neural networks seem to be something of little use, but curious. They know how to write music, process and generate images, highlight the main thing, voice text, and maintain a simple dialogue. But after the first delight, everyone will play enough, and the novelty will become a working commonplace in all areas. For example, several ways were selected specifically for the media to potentially use neural networks to solve real problems.

The article "Media of the Future: What Neural Networks Can Be Used in the Media Today " presents the results of a study by experts who, based on their many years of experience in online media, analyzed: what could simplify journalists' work, improve the quality of materials and increase business efficiency. Read more here.

Neural networks in Russia

In Russia, developments in the field of neural network programming are carried out by the largest Internet holdings, in particular, VK (formerly Mail.ru Group) and Yandex, using neural networks for image analysis and text processing in a search engine. The most famous examples were technologies from Microsoft, Google, IBM and Facebook, as well as startups MSQRD, PrismaOverview of ^[2]

2024

Russia has found a way to increase the efficiency of neural networks by 40%

Smart Engines scientists have found a way to increase the efficiency of neural networks. The method is based on a fundamentally new quantization scheme, thanks to which the speed of work increases by 40%. The results of the study were published in the journal Mathematics (Q1). Smart Engines announced this on April 25, 2024.

The development is already used in solving applied computer vision problems - for searching for objects and recognizing texts. It could also become an integral part of the next generation of unmanned autonomous systems, expanding the class of tasks that on-board computers can perform .

We are talking about the breakthrough of domestic scientists in the field of optimizing the execution of neural networks. As of April 2024, mainly neural networks are performed on specialized video cards, but not every computer is equipped with them. At the same time, any user device has a central processor, the world standard for which is the use of 8-bit neural networks. However, deep neural networks become more complex, containing hundreds of millions or more of coefficients, which require more computing power. This limits the use of central processors in artificial intelligence systems.

Smart Engines researchers solved this problem by offering a qualitative improvement to the 8-bit model - 4.6-bit networks. It works 40% faster than the 8-bit model, but is practically not inferior to it in quality due to the more efficient use of the features of the central processors of mobile devices.

To do this, the input data and coefficients of the model are quantized so that their products are placed in 8-bit registers. The summation of the results is done using a two-level system of 16- and 32-bit batteries to achieve maximum efficiency. As a result, on average, there are 4.6 bits of information per value.

Such a quantization scheme compares favorably with existing ones, since it allows you to flexibly set the bitness of the input data depending on the problem and does not bind to the powers of two. Therefore, this development provides noticeably higher quality recognition than, for example, 4-bit models.

Computer vision tasks should be solved on end devices - mobile phones, surveillance cameras, on-board computers of drones. All these tasks are characterized by low computing capabilities of devices and significant limitations on power consumption. And our development allows almost one and a half times to increase the possibilities of solving these problems. Classical networks in our recognition systems have already been replaced with 4.6-bit analogues, and we continue to work on more optimal schemes for quantizing and training neural networks, said Vladimir Arlazarov, CEO of Smart Engines, Doctor of Technical Sciences.

In all Smart Engines software products, "heavy" neural networks have been replaced by their 4.6-bit counterparts.

Russian scientists have improved the model of the diffusion neural network

Scientists from the Center for Artificial Intelligence and the Faculty of Computer Science of the Higher School of Economics, as well as the Institute of Artificial Intelligence AIRI and Sber AI have developed a diffusion model structure for which it is possible to set eight types of noise distribution. Instead of the classical structure of the Markov chain model and the application of normal distribution, scientists proposed a star-shaped model where it is possible to choose the type of distribution. This will help solve problems in different geometric spaces using diffusion models. This was announced on February 15, 2024 by representatives of the Higher School of Economics. Read more here.

Russian scientists have found a way to speed up the training of neural networks to navigate space

Researchers from the Higher School of Economics, NITU MISIS and AIRI have found a way to more effectively conduct reinforcement training for neural networks sharpened for orientation in space. With the help of the attention mechanism, the efficiency of the graph neural network increased by 15%. The results of the study are published in the journal IEEE Access. This was announced on January 23, 2024 by representatives of the Higher School of Economics. Read more here.

Russian scientists taught neural networks to recognize humor humanly

A group of scientists from the Faculty of Computer Science at the Higher School of Economics conducted a study of the ability of neural networks to recognize humor. It turned out that for more reliable recognition, the approach to creating data sets on which neural networks are trained should be changed. Such information was shared with TAdviser by representatives of the Higher School of Economics on January 11, 2024.

Freepik

As you know, voice assistants can only tell a ready-made anecdote, come up with their own or recognize the joking tone they are not able to. At the same time, users of voice assistants created on the basis of artificial intelligence technology want more humanity from them - the ability to recognize a joke and joke.

Since the mid-2000s, scientists have been engaged in recognizing humor as a classification task "funny is not funny," in the same frame datacets (a set of data) are collected and marked. A group of scientists from the HSE proposed to change the approaches to the formation of such datacets - to make them more diverse, and the data sets do not have to be very large.

As representatives of the Higher School of Economics explained, the task of recognizing humor is also difficult because there are no formal criteria for determining what is funny and what is not. Most existing datacets for learning and evaluating humor recognition models contain puns. Sarcasm and irony are even more complex, as is situational humor that requires knowledge of context or cultural code.

"We wanted to assess the tolerability and robustness of models trained on different datacets. Portability is how well a datacet-trained model with one type of humor defines a different type of humor. It was not at all obvious how the training would work, because humor is different, "said Pavel Braslavsky, associate professor at the Faculty of Computer Science at the Higher School of Economics.

Scientists tested the stability with "adversarial attacks" - attempts to force the neural network to see humor where it is not. The neural network received an unfunny, but formally similar to a humorous text - instead of a pun in the dialogue, a "wrong" consonant word was used. The smaller the network falls into such traps, the more stable it is.

The researchers trained the models on standard sensors for recognizing humor and on their mixtures. In addition, the models were tested with dialogue from Lewis Carroll's Alice in Wonderland, Charles Dickens' Antiquities Shops, Jerome K. Jerome's Three in a Boat, Not Counting a Dog, The Walking Dead, Friends, and a collection of ironic tweets.

It turned out that some models are retraining and consider everything ridiculous.

"We showed various models of Dickens' Antiquities Shop, which is a very sad story, and asked to assess what was happening. It turned out that some models believe that all dialogues from 19th-century literature are funny. And even more - everything that is too similar to the news of the XXI century is accepted as humor, "said Alexander Baranov, graduate student at the Faculty of Computer Science at the Higher School of Economics.

Models trained on puns are more often mistaken if in unfunny text one word is replaced with a consonant one. It also turned out that neural networks trained on small parts of different datasets recognize humor better than those trained on a large amount of the same type of data. The authors conclude that existing datacets are too narrow, the humor in each is very limited, and this reduces the quality of joke recognition.

The researchers proposed changing the approach to learning and evaluating models of humor recognition. We need new data sets, more diverse and close to ordinary conversations, natural communication. Large language models, such as ChatGPT, trained on huge amounts of data of different types, on average do a good job of recognizing humor, and scientists suggest that it is precisely the diversity of data they studied on.

"We are now only talking about binary recognition of humor: funny or not funny. It is very far from defining shades of humor, distinguishing between sarcasm and irony, recognizing situational, contextual humor. Our voice assistants have jokes so far "pinned with nails" and covered with filters that determine what kind of joke to issue depending on the user's words. This programming of responses feels unnatural. The demand for the greater humanity of artificial intelligence is absolutely understandable, but it will not be easy to satisfy it, "said Vladimir Knyazevsky, one of the authors of the study, a student at the Faculty of Computer Science at the Higher School of Economics.

The study was carried out as part of the project of the Scientific and Educational Laboratory of Models and Methods of Computational Pragmatics.^[3]

2022: Russian scientists presented a method for classifying photographs based on a quantum neural network

Russian physicists of the Laboratory of Quantum Information Technologies of the University of MISIS, the Russian Quantum Center and Moscow State University named after M.V. Lomonosov first presented a method of classification of photographs with high accuracy for 4 classes of images, based on the architecture of the quantum convolutional neural network (QCNN). Representatives of NUST MISIS reported this to TAdviser on November 24, 2022. Read more here.

2021

A neural network has been created in Russia that generates pictures according to the description in Russian

Sberbank on November 2, 2021 informed TAdviser about the creation of a neural network ruDALL-E, which is capable of creating images based on a text description in Russian. Read more here.

26% of entrepreneurs would trust artificial intelligence to communicate with customers

On October 13, 2021, Platforma (Big Data Platform) analysts were joined by the results of a study of the attitude of small and medium-sized businesses to artificial intelligence (AI) and big data. More than a third of respondents (36.8%) believe that the work of AI requires control and verification of results. But every fifth (22.1%) entrepreneur is already ready for October 2021 to replace a personal assistant with a smart program for fulfilling personal instructions. Read more here.

2020: Russian scientists taught artificial intelligence to "see" quantum advantages

Russian scientists from MIPT, FTIAN and ITMO have created a neural network that has learned to predict the behavior of a quantum system by "looking" at the scheme of this system. This was announced on January 16, 2020 by MIPT to TAdviser.

Such a neural network independently finds those solutions that are well suited for demonstrating quantum advantages. This will help researchers develop efficient quantum computers.

A large range of tasks of modern science is solved on the basis of quantum mechanical calculations. For example, chemical and biological: research on chemical reactions or the search for sustainable molecular structures for industry, medicine, pharmaceuticals and other fields.

Quantum computing is well suited for the exact solution of this kind of "quantum" problems, unlike classical ones, on the basis of which quantum problems are solved in most cases only cumbersome and approximate.

The process of creating quantum computing circuits is a laborious and expensive task. Not always the resulting devices show "quantum superiority" - they demonstrate the speed of information processing faster than a regular classical computer. Therefore, scientists would like to have a tool for predicting whether some scheme will have a quantum advantage or not.

One implementation of quantum computing is quantum wandering. Simplistically, you can imagine this method as moving a particle along a specific network made up of node points and connections between these nodes. Such networks form the circuit of a quantum system.

If the quantum movement of a particle - wandering - from one node of the network to another turns out to be faster than the classical one, then we can say that a device based on such a scheme shows a quantum advantage. Finding networks with a quantum advantage is an important task that experts in the field of quantum wandering are working on.

The idea of ​ ​ Alexei Melnikov, Leonid Fedichkin and Alexander Alojanets was to replace the expert with machine intelligence: to teach the computer to distinguish networks and give an answer to the question of which networks quantum wandering will give an advantage. That is, it makes sense to discover networks on the basis of which it makes sense to build a quantum computer.

The researchers took a neural network that "specialized" in image recognition. The network adjacency matrix and the number of the input and output node were supplied to the program input. At the output, the neural network gave an answer whether the quantum wandering between these nodes will be faster than the classical one.

It was not obvious that this approach would work, but it works, and we very successfully taught the computer to independently predict the quantum advantage in networks of complex structure, says Leonid Fedichkin, Associate Professor, Department of Theoretical Physics, Moscow Institute of Physics and Technology

The line between quantum and classical behavior of systems is often blurred. The highlight of our work was the creation of a special computer vision, with which we managed to see this line in the space of networks, explains Alexey Melnikov, researcher at ITMO

Researchers have created a tool to simplify the development of computational circuits based on quantum algorithms, the main applications of which should be biophotonics and materials science.

For example, quantum wanderings easily describe the excitation of photosensitive proteins such as rhodopsin or chlorophyll. A protein is, in a sense, a complex molecule that looks like a network. The task is to understand what will happen to an electron that has fallen into some point in the molecule, how it will move and what excitation it causes, in translation into formal language and there is a search for wandering time from one node of the network to another.

It is expected that the calculation of natural processes on quantum wanderings is easier to implement than on an architecture of qubits and gates, since wandering itself is a natural physical process.

2019: Promt unveils neural network translation solutions

In September 2019, Promt presented ready-made translation solutions using neural network and big data technologies - Promt Neural. The company calls them a new stage in the evolution of machine translation. Read more here.

2017

A neural network will be created in Moscow to recognize the readings of water meters by photo

The Department of Information Technologies (DIT) of Moscow has begun to create a system based on a neural network designed to read the readings of water meters directly from their photographs. It is planned to teach the neural network to recognize the readings of meters in the photo by the end of 2017. In order to train, she will have to process about 10 thousand such images.^[4]

The neural network created in Moscow will be able to read water flow readings from meter photographs. Photo: VistaNews.ru

The Moscow Mayor's Office on its website asked the residents of the city to help with the training of the neural network. To do this, they just need to upload pictures to the site, confirming then the correctness of the recognized numbers. Muscovites will be able to upload an unlimited number of photos, but they will have to follow a number of rules when photographing readings: the camera must be at a distance of no more than 15 cm from the meter; at least half of the image must be occupied by the counter image; one photo should not have two or more counters; You can take multiple shots of the same counter from different angles.

To report the meter readings, residents of the capital still have to enter data manually. Recognition of readings by photo is expected to take a matter of seconds and, as a result, will save Muscovites time.

After launching the trained neural network, photos will need to be uploaded to the applications HOUSING AND PUBLIC UTILITIES Moscow"," "Moscow State Services" or to the personal account on the site mos.ru. It is assumed that the network will be able to recognize numbers in meter photos regardless of lighting, shooting angle, camera capabilities and image quality.

Creation of cancer drugs Mail.Ru Group, Insilico Medicine and MIPT

On February 8, 2017, Mail.Ru Group, Insilico Medicine and MIPT reported the use of a neural network to create new drugs. It is assumed that this technology will help look for drugs in various fields - from oncology to cardiovascular diseases.

Russian scientists took as a basis the architecture of adversarial auto-encoders. Molecules with known therapeutic properties and effective concentration were used for training. Information about such a molecule was supplied to the input of the network. The network was configured to get exactly the same data at the output. It was composed of three structural elements - an encoder, decoder and discriminator - each of which performed its own specific role, "collaborating" with the other two.

The encoder, together with the decoder, trained to compress and then recover information about the original molecule, and the discriminator helped make the compressed representation more suitable for subsequent reconstruction. After the network was trained on many known molecules, the encoder, along with the discriminator, were "turned off," and the network, using a decoder, generated a description of the molecules itself.

Neural Network Architecture

According to experts, the effectiveness of training neural networks depends on the number of input data and the size of the network itself. On average, a good neural network can learn in a week. However, finding the optimal network architecture solution can take several months.

We made a neural network of a generative type, that is, able to create similar things on which she studied. We have trained a network model that is capable of creating new fingerprints with specified properties, - said one of the authors of the project, MIPT graduate student Andrei Kazennov.

In an experiment conducted by Mail.Ru Group, Insilico Medicine and MIPT, the network first studied on a variety of known molecules, then it generated a description of the molecules itself. To verify the network, a patent base in the field of cancer drugs was used. The task was to predict already known forms, but those that were not in the training sample. Most of the substances predicted by the neural network already have patents.

Neural networks in the world

2024: A neural network has been created that simulates the handwriting of a person

At the end of December 2023, specialists from the Mohamed bin Zayed University of Artificial Intelligence in the UAE (MBZUAI) announced the creation of a neural network capable of imitating human handwriting. The developers registered their technology with the United States Patent and Trademark Office (USPTO). Read more here.

2023: It turned out that neural networks are not able to learn like a human brain due to lack of sleep

At the end of November 2023, American specialists from the Massachusetts Institute of Technology (MIT) released the results of a study that studied the possibilities of deep neural networks in terms of imitating the human brain. Scientists have come to the conclusion that neural networks are not able to learn like humans due to lack of sleep.

The main task of deep neural networks is to simulate human cognitive abilities. However, due to the unprecedented complexity of the human brain, numerous difficulties arise. In particular, artificial intelligence uses predetermined parameters, which is why there are restrictions when processing unfamiliar or unfavorable scenarios.

Neural networks are not able to learn like humans due to lack of sleep

Research shows that while deep neural networks have made significant progress, they fail to fully mimic the human brain. In addition, such systems tend to overwrite existing data - a phenomenon known as catastrophic forgetting. This effect has a negative impact on AI learning speed. On the other hand, the human brain, having received new information, includes it in existing knowledge. The brain develops rational memory during rest: sleep allows you to form associations between objects and information that at first glance look unrelated to each other.

American researchers propose to include artificial sleep cycles in deep neural networks. It is assumed that this approach will help mitigate the impact of catastrophic forgetting and increase the effectiveness of training AI models. At the same time, scientists admit, despite progress, neural networks have a long way to go to achieve parity with human cognitive abilities.^[5]

2017: Recursive neural network managed to hack capcha

Scientists from the American company Vicarious have created an algorithm that decrypts capcha - the most common way to distinguish a person from a robot. Such an algorithm works on the basis of computer vision and a recursive cortical neural network and, according to the developers, can decipher capcha on many popular Internet platforms, including PayPal and Yahoo. The work is published in the journal Science.^[6]

CAPTCHA

Capcha (CAPTCHA, stands for Completely Automated Public Turing test to tell Computers and Humans Apart - a fully automated public Turing test that allows you to distinguish a person from a robot) is used to find out who is trying to use any service: a person or some program to automate actions on the Internet. Capcha is usually based on the task of, for example, distinguishing between "floating" letters, highlighting a word from the background, or marking photos that contain a specific object. To solve it, a person has enough knowledge about the world around him and basic skills (for example, reading). A computer, however, requires a huge amount of data to perform such a test. He can recognize any standard characters, but, for example, "floating" letters that are found for the first time - with difficulty. On the other hand, for a person, such a task does not pose a big problem; artificial intelligence, accordingly, must be maximally developed (compared to real, human intelligence) to solve it.

Recursive neural network

Scientists at Vicarious have been able to develop a neural network for decoding capcha, called the recursive cortical network (RCN). To create it, knowledge was used about the processing of visual information by a person, namely, about the effective separation of the object and background, even when they have a very similar structure. The created neural network is able to highlight the outline of the object (for example, letters) against the general background, even if part of the object is hidden after another.

Only about 26 thousand images were used to train the neural network. For comparison, the convolutional neural network (CNN) -based capch recognition algorithm requires several million.

To check the operation of the neural network, data from the open Google capch generator reCAPTCHA was used, the peculiarity of which, according to the developers, is their comparative ease of recognition for people and complexity for computers. In addition, Yahoo, PayPal and Botdetect captcha were used for verification.

Results of neural network testing

Capcha is considered solved if the computer managed to recognize it in one percent of cases. The neural network created by Vicarious was able to decrypt examples from reCAPTCHA with an accuracy of 66.6%. For comparison, a person can recognize the same combinations with an accuracy of 87%.

George et al. / Science 2017

Examples of captch used for training and the efficiency of the neural network at the level of words (third column) and letters (fourth column)

The algorithm also showed better (compared to other algorithms, the work of which is based on convolutional neural networks) efficiency in recognizing individual characters: up to 94.3%. For comparison, the efficiency of the convolutional neural network significantly decreases with increasing visual differences between the training and training samples.

George et al. / Science 2017

Efficiency of individual character recognition by a recursive neural network, or RCN, and a convolutional neural network, or CNN. y-axis - fraction of data difference from training and training samples

In general, the effective operation of the presented algorithm raises the question of the need to improve existing cybersecurity solutions and develop tools to protect user data from artificial intelligence.

How the structure of neural connections in the brain formed the basis of neural network algorithms

To understand the details of what a neural network is, together with mathematician Evgeny Putin, let's remember what neurons and synapses are and how their system works in living organisms. We learn how the structure of neural connections in the brain formed the basis of neural network algorithms, and what technological problems neural networks allow us to solve. About the pros and cons of working with neural networks; about how their training takes place, and what networks will be capable of in the near future.

Evgeny Putin, a graduate student at the Department of Computer Technologies at ITMO University, explores the problems of integrating the concept of feature selection into the mathematical apparatus of artificial neural networks.

Homo ex machina: Transfer consciousness to computer

If we talk about how realizable the idea of ​ ​ transferring a person's consciousness to a machine is, then it can be noted that technologies that in the future may form the basis for solving this problem are already developing today. The very task can be divided into two: the creation of a machine that could accommodate human consciousness, and the creation of technology that could copy this consciousness into this machine. What kind of machine should there be for it to be able to emulate human consciousness? How does the brain work, and what can be confirmation that solving the problem of transfer of consciousness is possible in principle? How do the mechanisms for transferring information to the brain work? What have scientists achieved in this direction in recent decades, and what can we do in this area in the near future?

Sergey Markov is a specialist in machine learning methods, founder of the XX2 VEK portal, author of one of the strongest Russian chess ^[7].

Graph neural networks: a fleeting trend or the future behind them

Main article: Graph neural networks

Graph neural networks are actively used in machine learning on graphs to solve local (classification of vertices, prediction of connections) and global (similarity of graphs, classification of graphs) problems. Local methods have many examples of applications in word processing, computer vision, and recommendation systems. Global methods, in turn, are used in approximating problems that are not effectively solved on modern computers (with the exception of the quantum computer of the future), and are used at the junction of computer and natural sciences to predict new properties and substances (this is relevant, for example, when creating new drugs).

The peak of popularity of graph neural networks reached in 2018, when they began to be used and showed high efficiency in various applications. The most famous example is the PinSage model in the service recommendation system. Since then, there Pinterest are more and more new applications of the technology in areas where previously existing methods were not able to effectively take into account in models of communication between objects. More. here

Notes