On the evolution of recurrent neural systems

The evolution of neural network architectures, first of the recurrent type and then with the use of attention technology, is considered. It shows how the approaches changed and how the developers’ experience was enriched. It is important that the neural networks themselves learn to understand the developers’ intentions and actually correct errors and flaws in technologies and architectures. Using new active elements instead of neurons expanded the scope of connectionist networks. It led to the emergence of new structures — Kolmogorov–Arnold Networks (KANs), which may become serious competitors to networks with artificial neurons.