Extended Data Fig. 8: External ramping input level determines robustness to distractors.
From: Attractor dynamics gate cortical information flow during decision-making
a, Schematics of distractor-free ramping RNN when an external ramping input of variable slope (purple) was delivered to the network. b, Proportion of distractor-evoked switching trials as a function of the time at which the distractor was delivered and of ramping input strength. Note that i) weaker ramping input induce more errors and ii) distractors delivered late in the delay are gated more efficiently. This suggests that elapsed time and amplitude of the ramping input were directly related: the gating of late distractors could be explained by the increased amplitude of the external ramping compared to earlier times in the delay. c, Assessing the contribution of external ramping input to temporal gating of distractors in the distractor-free ramping network. Left panel: To dissociate the impact on RNN dynamics of the external ramping input (which progressively increased during sample and delay epochs) from the effects of passage of time per se (that is regardless of the ramping), we clamped the external ramping input at the level it reached at the offset of the early-delay distractor (t = −1.2 s, black vertical line). Right panel: Proportion of switching trials for early- (gray) and late-delay (black) distractors: unlike the case examined in Fig. 4 (in which, for the same RNN, the external input kept ramping throughout the delay epoch), the proportion of switching trials evoked by late-delay distractor when the ramping input was clamped was similar to that of early-delay distractor. This indicates that temporal gating is regulated by the external ramp level and not by the passage of time per se. d, Relaxation time constants of distractor-free ramping network as a function of time at which the perturbation was delivered. In this test, the ramping input was set to the value it had reached at time ti in the delay (left panel), and, for each ramping input level, a perturbation of amplitude 0.2 a.u. along the Choice mode was delivered to the RNN during lick-left trials (middle panel). We then measured the time it took the network activity to drift back towards the lick left trajectory, that is the relaxation constant τ, by fitting a single exponential decay function \(f\left( t \right) = A\exp \left( { - \frac{t}{\tau }} \right) + B\) to the decaying trajectories. The relaxation time constant decreased as ramping slope increased (right panel), suggesting a deepening of the lick-left attraction basin with increased ramping (which corresponds, equivalently, to later times in the delay).
