Researchers from leading American institutes presented a new architecture of the Kolmogorov-Arnold Networks (KAN) neural network. It became an alternative to multilayer perceptrons (MLP), a principle developed in 1957.

Unlike the perceptron (MLP), which represents a simplified model of the biological neural network, where the center is a mathematical model of the perception of information by the brain, BLOOD is based on deep mathematical principles. So, on the approximation theorem or superposition theorem of Soviet mathematicians AN Kolmogorov and VI Arnold.

The researchers noted that KAN, unlike MLP, is able to process new information without catastrophic forgetting. The model is constantly updated without the need for any database or retraining.

KAN provides much better and more accurate answers than conventional models, but its training requires large computational power. The new architecture may create opportunities for further improvements in AI deep learning.