Discuss How do neural networks work?

[nikokik]

Neural networks have become part of everyday life today: they recognize faces in photos, translate texts, recommend movies, and even write music. But how do these "digital brains" cope with such tasks? Let's figure it out without going into mathematics, but through the prism of human thinking.

​

How is a neural network formed?​

The human brain processes information in layers: first he sees the contours, then the details, and then gathers them into a single image. The neural network works similarly. Each layer in it is responsible for its own level of analysis. The first layer can highlight the edges of objects, the next — textures, then shapes, and so on until the network gathers a complete picture. The more layers there are, the "deeper" the learning is — hence the term "deep learning". It is layering that allows neural networks to recognize complex patterns, such as emotions in text or tumors on X-rays.

Learning through Mistakes​

Imagine that you are learning to play darts. First, you hit anywhere except the target. But with each attempt, the brain corrects the movement of the hand, taking into account previous mistakes. The neural network trains similarly. After each prediction, she compares her answer with the correct one and calculates a "penalty" — an error. The stronger the error, the more actively it changes the internal settings in order to be more accurate next time. This process is repeated millions of times until the network learns to produce results with minimal errors.

Data​

Neural networks, like athletes, depend on the quality of "nutrition". The more diverse data they receive, the better they work. For example, if you're training a neural network to recognize dialects, it needs thousands of hours of audio recordings with different accents, intonations, and background noises. At the same time, the data must be "clean" — without errors in the markup. Otherwise, the network will learn the wrong lessons, like a child who remembers typos in a textbook.

Why do neural networks sometimes make mistakes?​

Even the most advanced neural network is not an omniscient oracle. Her mistakes are often related to a lack of data or a learning bias. For example, if the network was trained only on photos of light-skinned people, it may be worse at recognizing dark-skinned people. Or, if the data is dominated by images of cats on the couch, she won't immediately understand what the puss in boots from the cartoon looks like. In addition, neural networks do not cope well with tasks that require common sense. They can perfectly translate a text, but they can't understand a joke or sarcasm.

​

Conclusion​

In this article, I explained how a neural network works, how they learn and why they can make mistakes. If you still have any questions, please write in the comments.