Moscow Discoveries" at 14:46 on Radio Russia! Sort by color. Billiard-like Systems - Neural Network Tutor Training differential neural networks is a mathematical construct. When differential equations are coupled to the learning law, billiard-like systems are useful because they provide a large corpus of examples on which to train such a network. This allows us to solve the overfitting problem - to narrow it down to a narrow type of data. If you keep telling a neural network to recognize photographs of the same person, then it will only recognize him. Every color image is made up of pixels. Each pixel can be represented by a set of three numbers, that is, a decomposition into RGB colors: red, green and blue. Let's take a picture, put it on the table and from each point we will release a vector with specific coordinates. Instead of a flat image, a surface in three-dimensional space is obtained. Those pixels that correspond to the boundaries between objects, for example, an image of an animal against a nature background, become more visible. Program expert: Vladyslav Kivkalo - recipient of the Russian President's Prize in the field of science and innovation for young scientists for 2025, candidate for physical and mathematical sciences, associate professor at M.V. Lomonosov Moscow University. Presenter: Sergey Ivashko. Science journalist. Listen and subscribe to Radio Russia on Telegram MAX and download the official app of Radio Russia!

Moscow Discoveries" at 14:46 on Radio Russia! Sort by color. Billiard-like Systems - Neural Network Tutor Training differential neural networks is a mathematical construct. When differential equations are coupled to the learning law, billiard-like systems are useful because they provide a large corpus of examples on which to train such a network. This allows us to solve the overfitting problem - to narrow it down to a narrow type of data. If you keep telling a neural network to recognize photographs of the same person, then it will only recognize him. Every color image is made up of pixels. Each pixel can be represented by a set of three numbers, that is, a decomposition into RGB colors: red, green and blue. Let's take a picture, put it on the table and from each point we will release a vector with specific coordinates. Instead of a flat image, a surface in three-dimensional space is obtained. Those pixels that correspond to the boundaries between objects, for example, an image of an animal against a nature background, become more visible. Program expert: Vladyslav Kivkalo - recipient of the Russian President's Prize in the field of science and innovation for young scientists for 2025, candidate for physical and mathematical sciences, associate professor at M.V. Lomonosov Moscow University. Presenter: Sergey Ivashko. Science journalist. Listen and subscribe to Radio Russia on Telegram MAX and download the official app of Radio Russia!