Extended Data Fig. 9: Effects of multiple stimulations on behavior and ALM dynamics.

a, Task schematics with additional stimulation on lick-right trials. b, Probability to lick-right after single or multiple vS1 photostimulations. Error bars, Mean ± s.e.m. across sessions (n = 37 sessions). P values indicate statistical significance by a two-sided paired Student t-test. c, ALM neural activity projected on Choice mode on correct lick-right trials with additional stimulation during early delay (light-blue) or late delay (cyan); activity on trials without stimulation (red) or with single stimulation during sample epoch (blue) are shown for comparison. d, Same as in c, for additional stimulation during pre-sample epoch (dark blue). Data was computed based on activity of neurons recorded in left ALM (n = 1061 cells) of distractor-trained mice. To study the effect of additional stimuli on lick-right trials, we first compared a single stimulus during sample-epoch (‘Sample’) to additional identical stimuli during delay (‘Sample + Early Delay’, and ‘Sample + Late Delay’; a). We considered three possible outcomes: 1) Additional stimuli during delay would not affect performance because a lick-right instruction was already provided during sample epoch; 2) Additional stimuli provide additional evidence in favor of lick-right decision, which can be encoded by larger activity along the Choice mode in ALM. In this case, additional stimuli would increase the probability to lick-right regardless of their timing (early or late in the delay); 3) Additional stimuli on lick-right trials could ‘rescue’ neural dynamics on error trials, which would otherwise follow lick-left neural trajectory. Such rescue by an additional stimulus would consist of a switch from the erroneous lick-left trajectory to the correct lick-right trajectory. Stimulus-induced rescue can be expected to occur more often for early- rather than late-delay stimulation, because later in time the two attractors would be further apart from each other and thus less amenable to switching. Also, because in our attractor model selectivity (distance between attractors) is controlled by the non-specific ramping input, trials with multiple stimulations are not expected to result in higher selectivity at the end of the delay epoch. Our results supported the third possibility because: i) Additional stimulation during early delay (‘Sample + Early Delay’) but not late delay (‘Sample + Late Delay’) increased lick-right probability compared to single stimulation during sample (‘Sample’; b); ii) Choice mode trajectories were perturbed by the second stimulation during early or late delay, but then recovered to same level as the trajectory for single stimulation (c). These transient perturbations can be attributed to feedforward inhibition, as we observed that putative fast-spiking interneurons in ALM also responded to stimulation (Extended Data Fig. 2). The eventual recovery of the neural activity to the original lick-right trajectory, further supports the notion that lick-right trajectory was stabilized by attractor dynamics – which made the neural activity robust to transient perturbations. We also tested an experimental condition in which the additional stimulus was given during pre-sample epoch on lick-right trials (‘Pre-sample + Sample’, a). This increased the probability to lick right, and the Choice mode trajectory on these trials started earlier and ramped to a higher level compared to ‘Sample’ trajectory (b,d). The fact that both Pre-Sample stimulation and Early-Delay stimulation increased behavioral performance, but only Pre-Sample stimulation resulted in higher Choice mode activity by the end of the delay, is also consistent with a moving attractor model in which the fixed point corresponding to the lick-right attractor is moved in state-space by the ramping input. In this scenario, an earlier onset of the ramping input (for example during pre-sample) would result in larger separation between attractors by the end of the delay – which would be manifested by larger selectivity along the Choice mode (Supplementary Fig. 6). Taken together, these results showed that ALM dynamic: 1) was robust on lick-right trials, 2) was influenced by the time elapsed from the initiation of preparatory activity, and 3) did not encode the total number of stimuli. More broadly, these results suggest that similarly to frontal cortical regions in rats (Hanks, T. D. et al. 2015, Nature 520, 220–223), ALM activity in mice in this task encodes the categorical choice rather than graded strength of evidence.