Fig. 6: Behavioral manipulations trigger changes in coupling and latent representation.
From: Dynamical latent state computation in the male macaque posterior parietal cortex
a In separate experimental sessions, we manipulated sensory input (by altering the density of optic flow), motor output (by altering the sensitivity of the joystick), or latent state (by perturbing the monkey off his intended trajectory while traveling) (“Methods”). b These manipulations produced modest behavioral effects as shown by the ROC curves for a subset of experimental sessions, quantified as area under the curve (AUC, inset). c Coupling filters (shown for a representative subset of all directed pairs of five neurons) fit to data from baseline trials (gray) and manipulation trials (colored). d Top left: Cumulative distribution across all pairs of neurons of the correlation coefficient between coupling filters fit to baseline trials and manipulated trials. Black curve shows the baseline distribution computed using odd and even baseline trials. Gray curves show the null distribution constructed by shuffling the neuronal pairs. Shaded regions denote ±1 SEM estimated by bootstrapping (n = 25,805/20,798/22,506 neuron pairs for sensory/latent/motor manipulation). Top right: Cumulative distribution of the correlation coefficient between tuning function to sensory variables (the two sensory variables - linear and angular velocity - were concatenated) fit to baseline trials and manipulated trials across all neurons. Bottom left: Similar to top right, but computed using tuning to latent variables. Bottom right: Similar to top right, but computed using motor tuning. e Contribution of coupling to neuronal response, quantified as the improvement in the model log likelihood (LL) over uncoupled model, was comparable during baseline and manipulated trials. f Baseline-corrected stability to manipulated task (stability index) computed using cumulative distributions in d (“Methods”) for coupling filters, latent tuning, sensory tuning, and motor tuning under all three manipulations. Both sensory and motor tunings were robust to manipulations, whereas coupling and latent tunings were not. Source data are provided as a Source data file.
