Skip to main content
. 2021 Jul 22;15:678158. doi: 10.3389/fncom.2021.678158

Figure 11.

Figure 11

Training 2D GRUs. (top row) Phase portraits of target dynamical systems. Red solid lines represent 1-dimensional attractors. See main text for each system. (middle row) GRU dynamics learned from corresponding 29-step forecasting tasks. The prediction is an affine transformation of the hidden state. (bottom row) An example time series generated through closed-loop prediction of the trained GRU (denoted by a black trajectory). GRU fails to learn the ring attractor.