The Kyiv Academic University launched the course "Fundamentals of Neural Networks" as part of the Knowledge Rise project. The course was developed by Vitalii Tymchyshyn, Head of the Data and Machine Learning Laboratory at KAU, and Volodymyr Bezhuba, Senior Research Fellow of the same laboratory. Participants of the course were expected to create their first neural network, train it, evaluate the results, and test it in practice.

Neural Networks: How They Work Inside

The Kyiv Academic University launched the course "Fundamentals of Neural Networks" as part of the Knowledge Rise project.

The course was developed by Vitalii Tymchyshyn, Head of the Data and Machine Learning Laboratory at KAU, and Volodymyr Bezhuba, Senior Research Fellow of the same laboratory.

Participants of the course were expected to create their first neural network, train it, evaluate the results, and test it in practice.

Registration via Referral Link

Who is this course for? What knowledge will participants gain, and—most importantly—how can they apply it? Vitalii Tymchyshyn explains:

“Our idea was to create a basic course about neural networks, to explain what a network is on a conceptual level,” he says. “We will describe what neural networks are from a mathematical perspective, what network architecture means, and how to choose the best model design criteria.”

The researcher planned to demonstrate how to write code, share useful ideas for practical implementation, and discuss common problems encountered during training.
“During the course, we’ll create one or two simple neural networks using Python. Students will understand, figuratively speaking, how the ‘gears turn’ inside,” Tymchyshyn adds with a smile.

Why create neural networks at all?

“There are many problems without precise analytical solutions,” he explains. “For example, I can’t write a universal formula to determine whether a cat is in an image based on pixel brightness. But I can gather examples of images with and without cats and train a neural network to distinguish between them.”

Why ‘push’ information through the eye of a needle

According to the scientist, the course would be valuable for beginner programmers and anyone curious about how neural networks work.
Machine learning today is applied in nearly every field, and a course like this could become the first step toward mastering AI. After completion, students would be able to include in their CVs that they worked with the open-source machine learning library PyTorch, developed their own neural networks, and understood how they function.

The course consisted of four lectures.

The first introduced fundamental concepts and demonstrated how to create and train a neural network—such as one that recognizes handwritten digits.
The second lecture focused on neural network architectures: how neurons connect to one another and how networks are built from “layers.”

“The way and order in which layers are connected determines the architecture of the network,” Tymchyshyn explains. “I’ll describe interesting approaches where, for example, a neural network first reduces the dimensionality of input data and then expands it back to the original size. It’s fascinating because we essentially ‘push’ information through low-dimensional layers, like through the eye of a needle. If the information is restored at the output, that means its compressed form is sufficient to represent the data fully.”

This “compression,” according to Tymchyshyn, allows for reducing data size—an important feature for technical tasks like noise filtering, since only the most meaningful information passes through the “needle’s eye.”

Students also learned to solve gradient vanishing problems and apply skip connections—shortcuts that let information jump over one or more layers. These techniques are crucial for working with deep neural networks that are more difficult to train.

Testing dreams with business metrics

The third lecture was dedicated to training strategies, including overfitting and underfitting.
“We discussed how data quantity and training parameters affect performance and how to measure it correctly,” Tymchyshyn says.

Students learned how to assess network quality and when it performs well or poorly. “To evaluate performance properly, you must consider the business metric,” he notes. “Many people build networks for their own sake, without thinking about their real purpose.”

For instance, if a neural network analyzes medical scans for signs of cancer, a simple accuracy rate is not enough. “We must tolerate false positives—it’s better to double-check a healthy patient than miss an illness,” Tymchyshyn emphasizes. “A doctor can always confirm the result, but missing a disease could be fatal.”

The same applies in manufacturing—better to reject a good part than to let a defective one reach a car assembly line, where the consequences could be serious.

Thus, measuring success requires more than just accuracy; it must account for the real-world business context.

During the fourth lecture, the scientists planned to test neural networks on embedded devices. “We’ll try deploying them on Arduino boards used in robotics and drone manufacturing,” says Tymchyshyn. “This will be especially interesting for those pursuing defense technology.”

When asked why this project is personally meaningful, Tymchyshyn replied:

“I truly enjoy creating these lectures because they bring together the essential knowledge about neural networks in one concise course. I haven’t seen such short yet comprehensive programs before. Usually, they’re long and time-consuming. We want to make this one brief, engaging, and hands-on. It’s a challenge—to create, train, and test a network in just four lectures. But it’s definitely useful, and I’m sure students will love it. Highly recommended!”

Registration via Referral Link

By Svitlana Halata