Influence of heterogeneity and regulation on the stability–nonlinearity trade-off

Characterize how specific forms of structural heterogeneity in the recurrent weight matrix and dynamical regulation mechanisms such as automatic gain control influence the trade-off between dynamical stability and nonlinear computational power in recurrent neural networks used for reservoir computing.

Background

The paper demonstrates that introducing weakly coupled neuron subsets and applying global automatic gain control can suppress runaway excitation and improve performance across a wide range of excitatory/inhibitory balances.

Despite these advances, a principled understanding of how different heterogeneity patterns and regulation schemes jointly modulate stability versus nonlinear computational capability is not established, motivating a formal characterization of this trade-off.