Our graduate works in the organization, which performs technical diagnostics of pipeline transport objects. With time these steel structures are getting older.

There are two possible ways to deal with this problem: complete replacement (which is rather costly) or assessment of the technical condition (which is more beneficial). While assessing the damage, it is important to identify the defect in time and eliminate it, without allowing the total destruction of the structure.

The neural network that he had created managed to solve the problem of identifying and assessing the degree of danger of defects in the pipelines. The whole process was done totally autonomously without any human involvement, and the neural network evaluated data arrays and automatically detected defects, determining their types and parameters.

The set size for the diploma project was 50 thousand objects. In fact, the neural network was trained on an internal database of 120 million defects detected previously. This data was recorded during technical diagnostics applying modern methods of nondestructive testing.

"While working on the diploma, I realized that one should never be afraid to step back in the previous level (right into the "raw" data), and understand what the initial data was. Before the course, I thought diagnostic equipment could provide only 3 parameters as inputs for a neural network. It turned out to be much more.In time, I worked out what parameters have the greatest "influence" on the final metrics of a neural network operation and found 14 more parameters in the diagnostic ones. The final neural network for solving my task was built on them".

The neural network showed high accuracy in the process - the result exceeded all expectations! For comparison: when searching a person "misses" about 15% of defects , while the neural network does not exceed 1%. Also, the assessment of risk of defects done by people gave 20% error, whereas the neural network was able to reduce the error up to 5%. We were able to automate the process and improve the quality of work.

In the future,our graduate is planning to apply convolutional neural networks to solve the problem.

This student enrolled in the University of Artificial Intelligence to learn how to implement new technologies in the company. The past 25 years of programming experience have provided a good basis for understanding. But to master neural networks, you need to study Python, and learn the intricacies of all processes.

The company is engaged in automation of various processes at industrial enterprises. The idea was to create a neural network capable of detecting fire at an early stage by videotaping from surveillance cameras. In the future, such a "neural network" can be implemented as a supplement or instead of traditional fire alarm systems in offices, production facilities, warehouses, and open industrial sites.

Fire detectors usually trigger between 30 seconds and 2 minutes after a fire begins. During this time, the fire increases and it is much harder to put it out. A quick response solution to a fire in the first 20 to 30 seconds may prevent fires.

The neural network was tested on the database of 1000 photos with fire and smoke, 20 videos with the beginning of fires, 15 videos from office surveillance cameras. High accuracy of 90.8% was achieved, but the neural network could not recognize a small fire less than 1-2% of the image (500-1000 pixels). Then, it was decided to train the neural network on the frame difference to track small changes in the dynamics of the fire.

A year ago a strong need for a new product in the pharmaceutical market was noticed. The main problem was the absence of a unified accounting system for all pharmacy products. Each company has its own directories with copies, inaccuracies, and misprints. These create difficulties in search and accounting.

At that moment the reference directory consists of 170 000 items. It includes medicines and dietary supplements, as well as related products: candies, diapers, women's hygiene, cosmetics, and baby products.

The database makes a little more than 2 million records for network training.

Since the accounting system has different entries, marketers, and marketing companies cannot account for all sales and as a result do not receive full bonuses.

To automate the processes between large wholesale suppliers and pharmacies, there was created a system that could automatically find any spelling of the same goods in different nomenclature directories.

The test neural network was built based on the frequently used goods, but 8 months of writing the algorithm has not brought the desired results.

After meeting the founder of the University of Artificial Intelligence, the graduate realized that he could find the ideal mechanism.

He set a global goal to develop a system capable of automatically comparing the nomenclature reference book with a certain standard.

As a result of training, the accuracy of name recognition tends to indicators close to human recognition. Therefore, the program can soon become an excellent assistant in pharmacies and pharmacy chains, as well as in other areas in the pharmaceutical market such as pharmaceutical analytics.

This graduate has found a lot of opportunities to improve the system. For instance, to do text segmentation in the same way the human brain does.

This student applied to the program, having a solid base and all chances of success. In the past, he graduated from Moscow State University of Mathematics and Mechanics, department of mathematical forecasting methods under the guidance of Yuri Zhuravlyov. At MSU he got acquainted with the USSR developments in the field of methods now called Data Science.

At present, it is already forgotten that these methods were widely developed by the Soviet Union, including for combat systems. In his further career the student gained experience in C++ development, and implementation of corporate analytics systems on Oracle Applications platform, as well as business education in one of the leading European business schools.

Eventually, he created a product that allows to identify the disease of a garden plant by a series of photos. The result is an Android application that can be used by amateur gardeners, farmers and agricultural companies.

He took more than 3000 images for a database, and narrowed the choice to cucumbers and tomatoes. Every day for 2.5 months he spent on creating the neural network. In a training sample, the neuronet was able to recognize 17 types of diseases with an accuracy of 98%.

Applications in real gardens and fields require biotechnological research: a disease has dozens of symptoms and characteristics, subtle differences can only be detected with a microscope and research in the laboratory, after which the neural network can easily study them.

When the diploma thesis was finally finished, several similar products had already been created. Which means:

1. There is a competitive market, the product is in demand.
2. The idea is working, because it came to several developers at once.

After the termination of training the developer has joined the start-up company working in the field of application of Data Science to the problems of electronic marketing. The startup carries out development of AI systems and provides consulting services in its application on the data of the client. The diploma project worked fine as a portfolio and gave the student excellent career prospects.

A system administrator from Belarus has been working remotely in a Russian company for ten years. Realizing that prospects at work are limited, he started to look for a new topic for professional development.

"I've been keeping an eye on the sphere of artificial intelligence for a long time. When I saw a University course add, I made up my mind to learn how to build a neural network myself".

The goal was to put the project into practice at once. To identify its direction, the student met with a chief of the oncology department in a local hospital and found out that much help was required with mammograms.

There are 1.5 million women in the Republic who need to be examined annually, but there are not enough doctors. A program based on a neural network will help to screen the entire female population for breast cancer.

The main objective of the project was to come up with an algorithm which could detect 3 cohorts of the patients:

1.Those with no abnormal changes at all
2. Patients with abnormal changes in the form of a benign tumor
3. Patients with breast cancer.

The main base for the research were 340 samples of mammograms, but they were not enough.Having set the goal to increase the data size and attract attention to the idea, the student contacted the Department of Public Health in Belarus, and attended a meeting with the doctors. They expressed their strong view that the project would solve only a part of the problem and suggested highlighting the parts, representing tissue changes on mammograms.

"I have improved the neural network and managed to create two programs. The first program detects the presence or absence of cancer, leaving out benign tumours, with the accuracy 70 percent. The second one omits those patients who are healthy, with the accuracy 90 percent."

What I have now is one of the important blanks that can be used inside a scanner to detect if there is cancer. The program will be able to sort patients immediately: whether they have benign breast changes or not. And whether or not these changes are to be further examined.

The biggest problem, connected with the project implementation, is taking responsibility for making decisions. An error here can cost a human life. However, the program is not to replace a doctor, but to assist him.Specialists state that the neural network in its current form can already secure young doctors. The doctor won't make the mistake because the program will remind him. The specialist will be able to correct the program in case it suddenly makes a mistake.

"I'm currently in the process of obtaining data to improve the neural network. A bureaucratic problem interferes with it for the time being, but I will definitely solve this question".

A university lecturer decided to broaden his mind in the field of modern neural networks, and gain practical skills in their design. The object of his thesis was an important task for one of the divisions of the metallurgical plant.

The city metallurgical plant produces metal products. One of the intermediate links are steel billets of 300x360 cross section, on the end of which the mechanical method (a branding machine) puts a nine-digit mark.

One of the technological operations is loading these blanks into the furnace for heating before rolling. Loading is performed manually by the operator. Landing depends on the steel mark, which is hammered in the code. The operator must also compare this code with what is implemented in the database for each of the four furnaces for a specific customer. If he suddenly makes a mistake and puts the wrong billet in the furnace, the buyer will get the wrong billet and will charge the company with a fine, which is calculated at ten million rubles.

To help the operator with the purpose of duplicate control, a system is being created that provides automatic identification of the number of the cast blank.

Initially, the recognition task was performed by convolutional neural networks. However, with their help the task could not be solved completely, as recognition accuracy did not rise above 65%.

During the training the student learned about more powerful structures of neural networks capable of solving the task.Throughout the study, together with a team of like-minded people, the problem of determining the type of neural network that most accurately recognized the numbers of the mark was solved. The application of basic network architectures from the Segmentation models library were studied, namely Unet, FPN, Linknet, PSPNet. The resnet34 and seresnet34 networks were used as backbones. It also became possible to understand the installation, training and operation of FasterRCNN.

To create a dataset, a video camera was installed in the factory, which helped to assemble a network training base consisting of 10,000 images.

The study revealed that the best result from the networks of Segmentation models library is a combination of FPN with backbones seresnet34, and this result is comparable in quality with the result of FasterRCNN network.

As a result, in a test set of 1000 FPNs the network gave 90% accuracy, FRCNN - 92%. However, it should be noted that the FasterRCNN network is faster. It takes about 0.2 seconds to produce the result, while FPN spends 1 second on processing.

The plans are to improve recognition accuracy to 99%. There is a hypothesis that the best result will be obtained by combining these two networks.

The developer with 19 years of programming experience during his studies at the University of Artificial Intelligence created and tested a neural network which helps with document classification.

The main goals are to organize access to documents and improve the workflow in mailing lists. All documents stored in the workflow management system must be sorted by class. The process of document distribution to those professionals who work with them should become completely automated.

The task is to create a neural network that will be able to select from all documents contracts for the creation, and purchase of audiovisual works.

The neural network was trained on the database of 9,500 documents, tested and integrated into the django-based web service.

The test set to assess the quality of the neural network was obtained simply by taking 243 new documents from the system at the time of verification. The metrics of accuracy of completeness on the test set showed 99% - 240 documents out of 243 were correctly classified.

The 3 obtained classification errors in the test set were also studied in detail. In two cases out of three it was an error in data markup. And the remaining case represented a complex contract, which was difficult to classify manually.

This person came to the University to learn a promising profession and to enhance the skills of a specialist in artificial neural networks. Higher education in informatics and management in technical systems gave him only a general idea, without focusing on some particular skills. Realizing this, Ildar decided to study again.

In September 2018 he started to study JAVA and Python programming languages on his own. Then, having seen the perspective of neural networks, the student graduated from the University of Artificial Intelligence.

"The most important thing I got as a result of the training was an excellent skill. I'm planning to improve my skills and find a decent, high-paid job to enjoy doing what I really like and to help companies solve their problems using neural networks."

The diploma project helped me to thoroughly build a neural network. When solving the task of generating images of the watch, the hypothesis was tested whether GAN would be able to produce images of an absolutely original watch that has not yet been developed.

This required a carefully selected database of 11,000 images and 5 hours per day during 1.5 months to write the neural network and train it. The senior teacher's support gave the project new ideas.

As a result, it was possible to obtain images of new watch models in general. Clearly visible hands, dial, somewhere there are additional dials, and a date window. In the future, Ildar plans to finish the project to improve the result.

The head of the Information Systems Department in a satellite company STEKKOM LLC set the task to study the principles of neural network operation both for his own development and implementation in the company.

The student had a task to monitor different types of objects (people, animals, cars) with the camera. The lack of good communication was a problem - it was impossible to transmit the entire video stream.

It was decided to transmit not the entire volume of information, but a short semantic message. An inexpensive Nvidia Nano single-board computer with a GPU processor specialized for neural networks work was connected to the camera. It recognized a picture directly from the camera and transmitted short information about the found objects via satellite channel.

This technology can significantly reduce tracking costs. Full video streaming is expensive, and the tracking system for specific actions reduces costs.

Examples of commands for a neural network: to monitor an area, let the hosts in, to signal trespassing.

In creating the project, the Resnet-18 and GoogleNet models were used. The student took 1000 photos as a database (of his own and employees') and trained the network to determine who is who on the photos. Then the network learned to recognize faces in other images with an accuracy of 78%.

Most of the time (3 months) was spent not on training the network, but on learning the software that trains the network, as well as on forming the base.

This graduate has developed a long-standing interest in neural networks. At the University of Artificial Intelligence he conducted research on reducing the time and complexity of machine learning, and chose the task of face segmentation for this project. The basic level of programming proved to be sufficient for that.

What was done during the training:

I've invented my own algorithm which is not yet present on the market. It is usually necessary to "mark the base" for facial segmentation, i.e., to graphically mark the face in the photo.

This requires a lot of time and money to prepare the database.The algorithm allows segmentation of persons without preparing such a base. We used the usual base of 5000 photos of men and women, which is easy to get.

This algorithm greatly simplifies the introduction of neural networks for segmentation - and this is now one of the most popular tasks for business and research.

A year ago our student started a task of segmentation for an agricultural company. It was required to control the movement of animals in the agricultural complex. The traditional assistant library method did not work, so the developer started looking for a new solution. He learned that the most effective way to recognize objects is through neural networks and decided to go for training.

"My knowledge before studying at the University of Artificial Intelligence was fragmented, and my progress was slow. I realized that I needed a full immersion in the subject as well as support during the testing phase, so I enrolled for the course. As a result of the training I saw that neural networks are a modern method and a powerful tool for solving problems. I am glad that my old task has moved on and got a second chance for development".

In his graduation project the student tested animal recognition on video. The goal was to recognize a specific animal. The system had to precisely identify the name of a particular object.

Using a base of 5 hours of video and 10,000 pictures as a database, the student achieved 80-85% accuracy of recognition. The problem was that depending on the angle of the animal's movement, the neural network could not always recognize it correctly.

Now, the project is functioning in the test mode. The plans are to make the neural network more flexible so as not to depend on the video mode. The desired target is 90% accuracy.

There is a version that will help to improve the accuracy of the method of recognition of animal movement, as well as database expansion. This will improve the quality of the algorithm and may solve the problem with similar animals. It is also planned to consider using the Inceptionv4 or Inception-resnet network.

A teacher of a technical university, department of mathematical modeling took up the study of neural networks for self-development, and research purposes. Alexey is the most senior graduate of the University of AI: at the time of training he was over 60.

He has not been programming for many years - he once used the "old, good" Fortran. During his studies he had to pay attention to the Python language, and the basics of object-oriented programming.

"Studying programming almost from scratch was difficult, but very interesting at the same time. Thanks to the thoughtful course program, I was able to learn Python constructs step by step. The teachers - bright and charismatic - gave a lot of help and support. So, we were able to complete the training with the development of our project. And I plan to continue studying one of the new programs at the University of Artificial Intelligence".

The objective of the diploma was to create a neural network that could recognize the state of agricultural field plots, namely, to identify defective plots from the drone images.

When it comes to tens of hectares of land, there is a lack of manual labour done by specialists. To solve the problem the employees are trying to use remote sensing means. However, it is rather difficult to estimate defects from the obtained images. The neural network will be able to solve the problem on a scale where the team is unable to cope.

Reference images of qualitative and defective areas were offered to the developed network for training. It was able to differentiate what sites contained defects. The final neural network accuracy on the test set reached 71-83%. This is quite enough for the network to help in real work. It should be noted that the size of the training base was quite small - about 300 images, which required the implementation of augmentation. Increasing the size of the training set will significantly improve the accuracy of recognition in the future.

The plans are to develop the next stage, after which the improved architecture of the network will be able to classify areas by type of defect. According to the student this will help to improve the training set and adjust its macro parameters.