Paper ID: 2405.08825

Thermodynamic limit in learning period three

Yuichiro Terasaki, Kohei Nakajima

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 three data points? In this letter, we report the answer to be yes. Considering a random neural network in its thermodynamic limit, we show that under certain conditions, learning period three can embed attractors with all periods into the network as a bifurcation after learning. The associated universality is explained by a topological conjugacy between the trained network and the classical logistic map.

Submitted: May 12, 2024