Extended Data Fig. 8: Context-dependent decision-making state-space dynamics.

Here we reproduce figures akin to those presented in¹⁵ for the trained low-rank network used in Figs. 4 and 5. We generate 32 conditions corresponding to different combinations of context, signal A coherence and signal B coherence and then project condition-averaged trajectories either on the plane spanned by the recurrent connectivity vector m (which corresponds to the choice axis) and the input vector I^A, or on the m − I^B plane. Similarly to what was observed in¹⁵, signal A strength is encoded along the I^A axis, even when it is irrelevant (lower left corner), and signal B strength is encoded along the I^B axis, even when it is irrelevant (top right corner).