Physical Review Research, Journal Year: 2025, Volume and Issue: 7(2)
Published: May 5, 2025
A continuous one-dimensional map with period three includes all periods. This raises the following question: Can we obtain any types of periodic orbits solely by learning data points? In this paper, report answer to be yes. Considering a random neural network in its thermodynamic limit, first show that almost learned periods are unstable, and each has own characteristic attractors (which can even untrained ones). The latently acquired dynamics, which unstable within trained network, serve as foundation for diversity may lead emergence after learning. When interpolation is quadratic, universal post-learning bifurcation scenario appears, consistent topological conjugacy between classical logistic map. addition universality, explore specific properties certain networks, including singular behavior scale weight at infinite finite-size effects, symmetry three. Published American Physical Society 2025
Language: Английский