Neural networks: how artificial intelligence helps in business and life
Read the original article in the DTI Blog .
The 2013 Oxford Martin School said that 47% of all jobs could be automated over the next 20 years. The main driver of this process is the use of artificial intelligence, working with big data, as a more effective replacement for a person.
')
Machines are now capable of solving more and more processes for which people were previously responsible. In addition, they do it better and in many cases cheaper. About what this means for the labor market, German Gref spoke in July of this year, speaking to students of the Baltic Federal University. Kant:
Continuing to highlight the #technological future, the DTI team prepared everything you need to know for the first immersion in neural networks : how they work, why more and more companies prefer neural networks to live employees, and what potential for optimizing various processes this technology has.
The neural network is one of the ways to implement artificial intelligence (AI) .
In the development of AI, there is a vast area - machine learning . She is studying methods for constructing algorithms that can learn on their own. This is necessary if there is no clear solution to any problem. In this case, it is easier not to look for the right solution, but to create a mechanism that will come up with a method for its search.
help In many articles you can find the term "deep" - or "deep" - learning. By this is meant machine learning algorithms that use a lot of computing resources. In most cases, it can be simply understood as “neural networks”.
In order not to get lost in the concepts of "artificial intelligence", "machine learning" and "deep learning", we suggest looking at the visualization of their development:
# interesting There are two types of artificial intelligence (AI): weak (narrowly focused) and strong (general). Weak AI is designed to perform a narrow list of tasks. These are the voice assistants Siri and Google Assistant and all the other examples that we cite in this article. Strong AI, in turn, is able to perform any human task. At the moment, the implementation of a strong AI is impossible, it is a utopian idea.
The neural network models the work of the human nervous system, a feature of which is the ability to self-learn from previous experience. Thus, each time the system makes fewer errors.
Like our nervous system, a neural network consists of separate computational elements - neurons located on several layers. Data arriving at the input of the neural network is processed sequentially on each layer of the network. In addition, each neuron has certain parameters that can vary depending on the results obtained - this is the training of the network.
Suppose that the task of a neural network is to distinguish cats from dogs. To set up a neural network, a large array of signed images of cats and dogs is served. The neural network analyzes the signs (including lines, shapes, their size and color) on these pictures and builds a recognition model that minimizes the percentage of errors relative to the reference results.
The figure below shows the process of the neural network, whose task - to recognize the number of zip code, written by hand.
Despite the fact that neural networks are in the center of universal attention recently, it is one of the oldest machine learning algorithms. The first version of the formal neuron, the neural network cell, was proposed by Warren McCulloch and Walter Pitts in 1943.
And in 1958, Frank Rosenblatt developed the first neural network. Despite its simplicity, it could already distinguish, for example, objects in a two-dimensional space.
Mark I Perceptron - Rosenblatt Machine
The first successes attracted increased attention to technology, but then other machine learning algorithms began to show the best results, and the neural networks faded into the background. The next wave of interest occurred in the 1990s, after which almost no neural networks were heard until 2010.
Until 2010, there simply did not exist a database large enough to train a neural network in a high-quality way to solve certain tasks, mainly related to image recognition and classification. Therefore, neural networks are often mistaken: they confused a cat with a dog, or, even worse, a snapshot of a healthy organ with a snapshot of an organ affected by a tumor.
But in 2010, ImageNet appeared, containing 15 million images in 22 thousand categories. ImageNet was many times larger than the existing database of images and was accessible to any researcher. With such volumes of data, neural networks could be taught to make practically error-free decisions.
ImageNet size in comparison with other image databases existing in 2010
Prior to this, the development of neural networks was faced by another, no less significant, problem: the traditional teaching method was ineffective. Although the number of layers in the neural network plays an important role, the method of network training is also important. The previously used method of reverse encryption could effectively train only the last layers of the network. The learning process turned out to be too long for practical application, and the hidden layers of the deep neural networks did not function properly.
Results in solving this problem in 2006 were achieved by three independent groups of scientists. First, Jeffrey Hinton implemented network pre-training using the Boltzmann machine , teaching each layer separately. Secondly, Jan LeCahn suggested using a convolutional neural network to solve image recognition problems. Finally, Joshua Bengio developed a cascade auto-encoder that enabled all layers to be used in a deep neural network.
A team of researchers from the University of Nottingham has developed four machine learning algorithms to assess the risk of cardiovascular disease in patients. For training, data from 378 thousand British patients were used. The trained artificial intelligence determined the risk of cardiac diseases more effectively than real doctors. The accuracy of the algorithm is between 74 and 76.4 percent (the standard system of eight factors, developed by the American College of Cardiology, ensures accuracy of only 72.8%).
The Japanese insurance company Fukoku Mutual Life Insurance has signed a contract with IBM. According to him, 34 employees of the Japanese company will replace the system IBM Watson Explorer AI. The neural network will look at tens of thousands of medical certificates and take into account the number of visits to hospitals, surgery and other factors to determine the terms of insurance of customers. Fukoku Mutual Life Insurance is confident that using IBM Watson will increase productivity by 30% and pay off in two years.
Machine learning helps to recognize potential fraud in various areas of life. Such a tool uses, for example, PayPal - as part of the fight against money laundering, the company compares millions of transactions and detects suspicious ones among them. As a result, PayPal fraudulent transactions amount to a record low of 0.32%, while the standard in the financial sector is 1.32%.
Artificial intelligence has significantly improved the recommendation mechanisms in online stores and services. Algorithms based on machine learning analyze your behavior on the site and compare it with millions of other users. All in order to determine which product you are most likely to buy.
The recommendation engine provides Amazon with 35% of sales. The Brain algorithm, used by YouTube for recommending content, made it possible to achieve that almost 70% of the videos viewed on the site were found thanks to the recommendations (and not via links or subscriptions). The WSJ reported that the use of artificial intelligence for recommendations was one of the factors that influenced a 10-fold increase in the audience over the past five years.
The Yandex Data Factory algorithm is able to predict the impact of promotions on sales of goods. Analyzing the sales history, as well as the type and range of the store, the algorithm yielded 87% accurate (up to box) and 61% ultra-accurate (up to package) forecasts.
Neural networks that analyze natural language can be used to create chat bots , allowing customers to obtain the necessary information about the company's products. This will reduce the cost of call center teams. Such a robot is already working in the reception office of the Moscow Government and processes about 5% of requests. The bot is able to suggest, among other things, the location of the nearest MFC and the schedule for switching off hot water.
Albert, a full-cycle marketing platform that independently performs almost all operations, is also based on neural network technology . The lingerie manufacturer Cosabella, which uses it, eventually disbanded its own marketing department and fully trusted the platform.
Unmanned vehicles are a concept that most major concerns are working on, as well as technology companies (Google, Uber, Yandex and others) and startups, rely on neural networks in their work. Artificial intelligence is responsible for recognizing surrounding objects, be it another car, a pedestrian, or other obstacle.
So our world sees the neural network
The potential of artificial intelligence in this area is not limited to autopilot. A recent IBM survey showed that 74% of top managers in the automotive industry expect smart cars to appear on the roads by 2025. Such cars, integrated into the Internet of Things (see our previous longrid ), will collect information about the preferences of passengers and automatically adjust the cabin temperature, the radio volume, the position of the seats and other parameters. In addition to piloting, the system will also inform about any problems (and even try to solve them) and the situation on the road.
The neural network, developed by Mark Waller of Shanghai University, specializes in the development of synthetic molecules . The algorithm consisted of a six-step synthesis of a benzopyran sulfonamide derivative (necessary for Alzheimer's treatment) in just 5.4 seconds.
Yandex Data Factory tools help in steel smelting : scrap metal used in steel production is often heterogeneous in composition. In order for steel to conform to standards, when smelting it, one must always take into account the specifics of scrap and introduce special additives. This is usually done by specially trained technologists. But, since such information gathers a lot of information about the incoming raw materials, the additives used and the result, this information is more efficiently able to process the neural network. According to Yandex, the introduction of neural networks can reduce the cost of expensive ferroalloys by 5%.
Similarly, a neural network can assist in the processing of glass. Now it is an unprofitable, albeit useful, business that needs government subsidies. The use of machine learning technologies will significantly reduce costs.
Microsoft engineers together with scientists from ICRISAT use artificial intelligence to determine the optimal planting time in India . An application using Microsoft Cortana Intelligence Suite also monitors the condition of the soil and selects the necessary fertilizers. Initially, only 175 farmers from 7 villages participated in the program. They started sowing only after a corresponding SMS notification. As a result, they harvested 30–40% more than usual.
Last year, applications that use neural networks to process photos and videos instantly became popular: MSQRD from Belarusian developers (later bought out the service by Facebook), and Russian Prisma and Mlvch. Another service, Algorithmia, paints black and white photographs.
Yandex is successfully experimenting with music: the company's neural networks have already recorded two albums: in the style of Nirvana and “Civil Defense” . And the music, written by the neural network under the classic composer Alexander Scriabin, was performed by the chamber orchestra, which makes it necessary to re-think about the question whether the robot can compose a symphony. The neural network created by Sony employees was inspired by Bach.
The Japanese algorithm wrote the book “The Day When the Computer Wrote the Novel”. Despite the fact that people helped the inexperienced writer with the characters and plot lines, the computer did a great job - as a result, one of his works passed the qualifying round of the prestigious literary prize. Neural networks also wrote sequels to Harry Potter and Game of Thrones .
In 2015, the AlphaGo neural network, developed by the Google DeepMind team, became the first program to win a professional go player. And in May of this year, the program beat the strongest go player in the world , Ke Tse. This was a breakthrough, because for a long time it was believed that computers did not have the intuition necessary for playing go.
A development team from the University of Technology Sydney introduced drones to patrol beaches. The main task of drones will be the search for sharks in coastal waters and warning people on the beaches . Video data analysis is produced by neural networks, which significantly affected the results: the developers claim that the probability of detecting and identifying sharks is up to 90%, while the operator watching video from the drones successfully recognizes sharks only 20-30% of cases.
Australia ranks second in the world after the United States in the number of cases of shark attacks on people. In 2016, 26 cases of shark attacks were recorded in this country, two of which resulted in the death of people.
In 2014, Kaspersky Lab reported that their antivirus registers 325 thousand new infected files daily. At the same time, the research of the Deep Instinct company showed that the new versions of viruses practically do not differ from the previous ones - the change ranges from 2% to 10%. The self-learning model developed by Deep Instinct, based on this information, can detect infected files with high accuracy.
Neural networks are also able to look for certain patterns in how information is stored in cloud services, and to report detected anomalies that could lead to a security breach.
In 2016, 65-year-old NVIDIA engineer Robert Bond ran into a problem: neighboring cats regularly visited his site and left traces of their presence, which irritated his wife, who works in the garden. Bond immediately compartment too unfriendly idea to build a trap for uninvited guests. Instead, he decided to write an algorithm that would automatically turn on garden sprinklers as the cats approach.
Robert was faced with the task of identifying cats in the video stream coming from an external camera. For this, he used a system based on the popular Caffe neural network. Each time the camera observed a change in the situation on the site, she took seven pictures and transmitted them to the neural network. After that, the neural network had to determine whether a cat is present in the frame, and, in the case of an affirmative answer, turn on sprinklers.
The image from the camera in the courtyard of Bond
Before starting work, the neural network was trained: Bond “fed” 300 different photos of cats to her. Analyzing these photos, the neural network learned to recognize animals. But this was not enough: she correctly identified cats only in 30% of cases and took the shadow of Bond for a cat, as a result of which he turned out to be wet.
The neural network earned better after additional training on more photos. However, Bond warns that the neural network can be trained too much, in which case it will have an unrealistic stereotype - for example, if all the pictures used for training are taken from one angle, then artificial intelligence may not recognize the same cat from another angle. Therefore, it is extremely important to competently select a training data series.
After a while, cats that had not learned from photographs, but in their own skin, stopped visiting the Bond site.
Neural networks, a technology from the middle of the last century, are now changing the work of entire industries. The reaction of society is ambiguous: some of the capabilities of neural networks are fascinated, while others are forced to doubt their usefulness as specialists.
However, not everywhere where machine learning comes, it displaces people. If the neural network 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 Deep Blue supercomputer won chess against Garry Kasparov in 1997, but people from chess have not gone away, and eminent grandmasters still fall on the covers of glossy magazines.
Co-operation with machines will be much more beneficial than confrontation. Therefore, we have compiled a list of materials in the public domain that will help you to continue acquaintance with neural networks:
The 2013 Oxford Martin School said that 47% of all jobs could be automated over the next 20 years. The main driver of this process is the use of artificial intelligence, working with big data, as a more effective replacement for a person.
')
Machines are now capable of solving more and more processes for which people were previously responsible. In addition, they do it better and in many cases cheaper. About what this means for the labor market, German Gref spoke in July of this year, speaking to students of the Baltic Federal University. Kant:
“ We stop hiring lawyers who don’t know what to do with the neural network. <...> You are the students of yesterday. Comrades lawyers, forget your profession. Last year, the 450 lawyers who are preparing lawsuits are gone, were cut. Our neural net is preparing claims better than lawyers trained by the Baltic Federal University. We will definitely not take them to work. ”
Continuing to highlight the #technological future, the DTI team prepared everything you need to know for the first immersion in neural networks : how they work, why more and more companies prefer neural networks to live employees, and what potential for optimizing various processes this technology has.
Artificial Intelligence, Machine Learning, and Neural Networks: What Is the Difference?
The neural network is one of the ways to implement artificial intelligence (AI) .
In the development of AI, there is a vast area - machine learning . She is studying methods for constructing algorithms that can learn on their own. This is necessary if there is no clear solution to any problem. In this case, it is easier not to look for the right solution, but to create a mechanism that will come up with a method for its search.
help In many articles you can find the term "deep" - or "deep" - learning. By this is meant machine learning algorithms that use a lot of computing resources. In most cases, it can be simply understood as “neural networks”.
In order not to get lost in the concepts of "artificial intelligence", "machine learning" and "deep learning", we suggest looking at the visualization of their development:
# interesting There are two types of artificial intelligence (AI): weak (narrowly focused) and strong (general). Weak AI is designed to perform a narrow list of tasks. These are the voice assistants Siri and Google Assistant and all the other examples that we cite in this article. Strong AI, in turn, is able to perform any human task. At the moment, the implementation of a strong AI is impossible, it is a utopian idea.
How is the neural network
The neural network models the work of the human nervous system, a feature of which is the ability to self-learn from previous experience. Thus, each time the system makes fewer errors.
Like our nervous system, a neural network consists of separate computational elements - neurons located on several layers. Data arriving at the input of the neural network is processed sequentially on each layer of the network. In addition, each neuron has certain parameters that can vary depending on the results obtained - this is the training of the network.
Suppose that the task of a neural network is to distinguish cats from dogs. To set up a neural network, a large array of signed images of cats and dogs is served. The neural network analyzes the signs (including lines, shapes, their size and color) on these pictures and builds a recognition model that minimizes the percentage of errors relative to the reference results.
The figure below shows the process of the neural network, whose task - to recognize the number of zip code, written by hand.
Neural Network History
Despite the fact that neural networks are in the center of universal attention recently, it is one of the oldest machine learning algorithms. The first version of the formal neuron, the neural network cell, was proposed by Warren McCulloch and Walter Pitts in 1943.
And in 1958, Frank Rosenblatt developed the first neural network. Despite its simplicity, it could already distinguish, for example, objects in a two-dimensional space.
Mark I Perceptron - Rosenblatt Machine
The first successes attracted increased attention to technology, but then other machine learning algorithms began to show the best results, and the neural networks faded into the background. The next wave of interest occurred in the 1990s, after which almost no neural networks were heard until 2010.
Why neural networks are again popular
Until 2010, there simply did not exist a database large enough to train a neural network in a high-quality way to solve certain tasks, mainly related to image recognition and classification. Therefore, neural networks are often mistaken: they confused a cat with a dog, or, even worse, a snapshot of a healthy organ with a snapshot of an organ affected by a tumor.
But in 2010, ImageNet appeared, containing 15 million images in 22 thousand categories. ImageNet was many times larger than the existing database of images and was accessible to any researcher. With such volumes of data, neural networks could be taught to make practically error-free decisions.
ImageNet size in comparison with other image databases existing in 2010
Prior to this, the development of neural networks was faced by another, no less significant, problem: the traditional teaching method was ineffective. Although the number of layers in the neural network plays an important role, the method of network training is also important. The previously used method of reverse encryption could effectively train only the last layers of the network. The learning process turned out to be too long for practical application, and the hidden layers of the deep neural networks did not function properly.
Results in solving this problem in 2006 were achieved by three independent groups of scientists. First, Jeffrey Hinton implemented network pre-training using the Boltzmann machine , teaching each layer separately. Secondly, Jan LeCahn suggested using a convolutional neural network to solve image recognition problems. Finally, Joshua Bengio developed a cascade auto-encoder that enabled all layers to be used in a deep neural network.
Examples of successful use of neural networks in business
The medicine
A team of researchers from the University of Nottingham has developed four machine learning algorithms to assess the risk of cardiovascular disease in patients. For training, data from 378 thousand British patients were used. The trained artificial intelligence determined the risk of cardiac diseases more effectively than real doctors. The accuracy of the algorithm is between 74 and 76.4 percent (the standard system of eight factors, developed by the American College of Cardiology, ensures accuracy of only 72.8%).
Finance
The Japanese insurance company Fukoku Mutual Life Insurance has signed a contract with IBM. According to him, 34 employees of the Japanese company will replace the system IBM Watson Explorer AI. The neural network will look at tens of thousands of medical certificates and take into account the number of visits to hospitals, surgery and other factors to determine the terms of insurance of customers. Fukoku Mutual Life Insurance is confident that using IBM Watson will increase productivity by 30% and pay off in two years.
Machine learning helps to recognize potential fraud in various areas of life. Such a tool uses, for example, PayPal - as part of the fight against money laundering, the company compares millions of transactions and detects suspicious ones among them. As a result, PayPal fraudulent transactions amount to a record low of 0.32%, while the standard in the financial sector is 1.32%.
Commerce
Artificial intelligence has significantly improved the recommendation mechanisms in online stores and services. Algorithms based on machine learning analyze your behavior on the site and compare it with millions of other users. All in order to determine which product you are most likely to buy.
The recommendation engine provides Amazon with 35% of sales. The Brain algorithm, used by YouTube for recommending content, made it possible to achieve that almost 70% of the videos viewed on the site were found thanks to the recommendations (and not via links or subscriptions). The WSJ reported that the use of artificial intelligence for recommendations was one of the factors that influenced a 10-fold increase in the audience over the past five years.
The Yandex Data Factory algorithm is able to predict the impact of promotions on sales of goods. Analyzing the sales history, as well as the type and range of the store, the algorithm yielded 87% accurate (up to box) and 61% ultra-accurate (up to package) forecasts.
Neural networks that analyze natural language can be used to create chat bots , allowing customers to obtain the necessary information about the company's products. This will reduce the cost of call center teams. Such a robot is already working in the reception office of the Moscow Government and processes about 5% of requests. The bot is able to suggest, among other things, the location of the nearest MFC and the schedule for switching off hot water.
Albert, a full-cycle marketing platform that independently performs almost all operations, is also based on neural network technology . The lingerie manufacturer Cosabella, which uses it, eventually disbanded its own marketing department and fully trusted the platform.
Transport
Unmanned vehicles are a concept that most major concerns are working on, as well as technology companies (Google, Uber, Yandex and others) and startups, rely on neural networks in their work. Artificial intelligence is responsible for recognizing surrounding objects, be it another car, a pedestrian, or other obstacle.
So our world sees the neural network
The potential of artificial intelligence in this area is not limited to autopilot. A recent IBM survey showed that 74% of top managers in the automotive industry expect smart cars to appear on the roads by 2025. Such cars, integrated into the Internet of Things (see our previous longrid ), will collect information about the preferences of passengers and automatically adjust the cabin temperature, the radio volume, the position of the seats and other parameters. In addition to piloting, the system will also inform about any problems (and even try to solve them) and the situation on the road.
Industry
The neural network, developed by Mark Waller of Shanghai University, specializes in the development of synthetic molecules . The algorithm consisted of a six-step synthesis of a benzopyran sulfonamide derivative (necessary for Alzheimer's treatment) in just 5.4 seconds.
Yandex Data Factory tools help in steel smelting : scrap metal used in steel production is often heterogeneous in composition. In order for steel to conform to standards, when smelting it, one must always take into account the specifics of scrap and introduce special additives. This is usually done by specially trained technologists. But, since such information gathers a lot of information about the incoming raw materials, the additives used and the result, this information is more efficiently able to process the neural network. According to Yandex, the introduction of neural networks can reduce the cost of expensive ferroalloys by 5%.
Similarly, a neural network can assist in the processing of glass. Now it is an unprofitable, albeit useful, business that needs government subsidies. The use of machine learning technologies will significantly reduce costs.
Agriculture
Microsoft engineers together with scientists from ICRISAT use artificial intelligence to determine the optimal planting time in India . An application using Microsoft Cortana Intelligence Suite also monitors the condition of the soil and selects the necessary fertilizers. Initially, only 175 farmers from 7 villages participated in the program. They started sowing only after a corresponding SMS notification. As a result, they harvested 30–40% more than usual.
Entertainment and art
Last year, applications that use neural networks to process photos and videos instantly became popular: MSQRD from Belarusian developers (later bought out the service by Facebook), and Russian Prisma and Mlvch. Another service, Algorithmia, paints black and white photographs.
Yandex is successfully experimenting with music: the company's neural networks have already recorded two albums: in the style of Nirvana and “Civil Defense” . And the music, written by the neural network under the classic composer Alexander Scriabin, was performed by the chamber orchestra, which makes it necessary to re-think about the question whether the robot can compose a symphony. The neural network created by Sony employees was inspired by Bach.
The Japanese algorithm wrote the book “The Day When the Computer Wrote the Novel”. Despite the fact that people helped the inexperienced writer with the characters and plot lines, the computer did a great job - as a result, one of his works passed the qualifying round of the prestigious literary prize. Neural networks also wrote sequels to Harry Potter and Game of Thrones .
In 2015, the AlphaGo neural network, developed by the Google DeepMind team, became the first program to win a professional go player. And in May of this year, the program beat the strongest go player in the world , Ke Tse. This was a breakthrough, because for a long time it was believed that computers did not have the intuition necessary for playing go.
Security
A development team from the University of Technology Sydney introduced drones to patrol beaches. The main task of drones will be the search for sharks in coastal waters and warning people on the beaches . Video data analysis is produced by neural networks, which significantly affected the results: the developers claim that the probability of detecting and identifying sharks is up to 90%, while the operator watching video from the drones successfully recognizes sharks only 20-30% of cases.
Australia ranks second in the world after the United States in the number of cases of shark attacks on people. In 2016, 26 cases of shark attacks were recorded in this country, two of which resulted in the death of people.
In 2014, Kaspersky Lab reported that their antivirus registers 325 thousand new infected files daily. At the same time, the research of the Deep Instinct company showed that the new versions of viruses practically do not differ from the previous ones - the change ranges from 2% to 10%. The self-learning model developed by Deep Instinct, based on this information, can detect infected files with high accuracy.
Neural networks are also able to look for certain patterns in how information is stored in cloud services, and to report detected anomalies that could lead to a security breach.
Bonus: neural networks guard our lawn
In 2016, 65-year-old NVIDIA engineer Robert Bond ran into a problem: neighboring cats regularly visited his site and left traces of their presence, which irritated his wife, who works in the garden. Bond immediately compartment too unfriendly idea to build a trap for uninvited guests. Instead, he decided to write an algorithm that would automatically turn on garden sprinklers as the cats approach.
Robert was faced with the task of identifying cats in the video stream coming from an external camera. For this, he used a system based on the popular Caffe neural network. Each time the camera observed a change in the situation on the site, she took seven pictures and transmitted them to the neural network. After that, the neural network had to determine whether a cat is present in the frame, and, in the case of an affirmative answer, turn on sprinklers.
The image from the camera in the courtyard of Bond
Before starting work, the neural network was trained: Bond “fed” 300 different photos of cats to her. Analyzing these photos, the neural network learned to recognize animals. But this was not enough: she correctly identified cats only in 30% of cases and took the shadow of Bond for a cat, as a result of which he turned out to be wet.
The neural network earned better after additional training on more photos. However, Bond warns that the neural network can be trained too much, in which case it will have an unrealistic stereotype - for example, if all the pictures used for training are taken from one angle, then artificial intelligence may not recognize the same cat from another angle. Therefore, it is extremely important to competently select a training data series.
After a while, cats that had not learned from photographs, but in their own skin, stopped visiting the Bond site.
Conclusion
Neural networks, a technology from the middle of the last century, are now changing the work of entire industries. The reaction of society is ambiguous: some of the capabilities of neural networks are fascinated, while others are forced to doubt their usefulness as specialists.
However, not everywhere where machine learning comes, it displaces people. If the neural network 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 Deep Blue supercomputer won chess against Garry Kasparov in 1997, but people from chess have not gone away, and eminent grandmasters still fall on the covers of glossy magazines.
Co-operation with machines will be much more beneficial than confrontation. Therefore, we have compiled a list of materials in the public domain that will help you to continue acquaintance with neural networks:
- The course “Machine Learning” from the School of Data Analysis (ShAD) of Yandex
- Google Deep Learning Course
- The book "Mathematical teaching methods for precedents (theory of machine learning)" K. V. Vorontsova
- The book "Neurocomputer Engineering: Theory and Practice" F. Wassermen
- Movidius Neural Compute Stick - neural network the size of a USB flash drive
- A selection of machine learning frameworks
Source: https://habr.com/ru/post/337870/
All Articles