Skip to main content
. 2023 Oct 28;14:6869. doi: 10.1038/s41467-023-42609-4

Fig. 8. Enhanced robustness of the choice code in ALM by PPC.

Fig. 8

a Schematic of attractor dynamics along the choice axis during the delay. b Schematic of the RNN architecture trained to mimic the activity of ALM neurons. A stimulus pulse was delivered to the RNN instructing the right trial type. In perturbation trials, a distractor was delivered to the trained RNN. c Example RNN units (right) reproducing the activity of ALM neurons (left) trained either with PPC-ALM ( + PPC) or vS1-ALM (+vS1) external activity. d Example population activity of ALM units projected to the choice axis. Red and blue trajectories indicate average choice activity for right and left trial types without a distractor, respectively. Dashed gray and solid black trajectories indicate individual perturbation trials that did and did not switch decision, respectively. Shaded green rectangle denotes the distraction epoch. e Left. Pairwise comparison of fractions of decision-switching trials between RNNs trained with PPC and vS1 (n = 21 RNNs). Red cross indicates mean ± SEM. Right. Fractions of decision-switching trials between RNNs trained with PPC and vS1 (n = 21 RNNs, ***P < 0.001, one-tailed bootstrap). Error bars indicate mean ± SEM. f Left. Schematic of an RNN similar to (b) but with random ablation of connections between ALM and external units from PPC. Right. Population activity of an example RNN trained with PPC-ALM external units after 40% ablation. g Left. Pairwise comparison of fractions of decision-switching trials between RNNs with and without 40% ablation (n = 21 RNNs). Red cross indicates mean ± SEM. Right. Fractions of decision-switching trials as a function of percentage of connections ablated (n = 21 RNNs, **P = 0.005, one-way ANOVA). Error bars indicate mean ± SEM. h Left. Same as (d) but with an example RNN trained with the same ALM units but using reconstructed PPC-ALM external activity from a naive session. Right. Fractions of decision-switching trials between RNNs across learning (+PPC: ***P < 0.001; +vS1: n.s., P = 0.46, naive: n = 20; expert: n = 21 RNNs, one-tailed bootstrap). Error bars indicate mean ± SEM. i Schematic illustrating how PPC influences the choice-related attractor dynamics in ALM.