While it possesses many desirable structural properties, the estimation of high-dimensional MI from samples suffers from the curse of dimensionality.

Tostudy thishypothesis, wepropose GatedL0 Regularized Dynamics (GateL0RD), a novel recurrent architecture that incorporates the inductive bias to maintain stable, sparsely changing latent states.

"collapsing" any uncertainty, but when an arm is passive, no observation is made,

To improve the forecasting capabilities, we extend this classical model by conditionally linear state-to-switch dynamics, while leaving thepartial tractability oftheconditional Gaussian linear part intact.