Latching dynamics for neural networks with a sigmoidal nonlinearity

This article is intended to clarify the relationship between a kind of dynamical sequential retrieving of learned states (‘latching dynamics’) which have been observed in neural networks with sigmoidal nonlinearity, and the same dynamics analyzed by Köksal-Ersöz et al. (2020) on a simpler model which allows to give proofs and criteria for the occurrence of this behavior, but which is a singular approximation of the original model. • Latching dynamics in a neural network produces sequences of memorized states. • A mathematical analysis for a system with sigmoid nonlinearity is provided. • This extends previous results obtained on a simplified model. • Results highlight similarities and differences between the two models.