Noise induced escape in one-population and two-population stochastic neural networks with internal states

In the present paper, the escapes from the basins of fixed points induced by intrinsic noise are investigated in both one- and two-population stochastic hybrid neural networks. In the weak noise limit, the quasipotentials are computed through the application of WKB approximation to the original hybrid system and the results of quasi-steady-state (QSS) diffusion approximation. It is seen that the two results are consistent with each other within the neighborhood of a fixed point and an obvious discrepancy arises in the other area, of which the reason is then explored and revealed. Furthermore, the relationship between the fluctuational paths and the relaxational ones is analyzed, based on which some specific results for the hybrid system is obtained. Besides, for the two-population model, the phenomenon of saddle point avoidance is investigated by using both theoretical and numerical methods. Finally, the topological structure of Lagrangian manifold is analyzed, and its particular features and something analogous to the stochastic differential equation are found according to the accuracy of QSS within the vicinity of the saddle point.