RNN Forward Approach
An approach to model a Recurrent Neural Network based on a State Space Model where the current state is being used to calculate the next state. So it takes at least one more time step to see the effect in the observations.
s_{t+1}=f\left(s_t, u_t\right) \\ y_t=g\left(s_t\right) \end{gathered}$$ We can translate this [[Stetigkeit|continous]] model to a [[Neural Network]] structure via universal approximation: ![[Bildschirmfoto 2023-04-11 um 23.15.26.png]] ## Architecture - Allows to include $y_t$ as input - Shift along the time labeling ![[Bildschirmfoto 2023-04-11 um 23.19.26.png]]