Extended Data Fig. 8: The distribution of commitment times inferred from MMDDM does not match the distribution of start time of peri-movement kernels.
From: Transitions in dynamical regime and neural mode during perceptual decisions
a, Separately for each choice-selective neuron (N = 4605), peri-movement kernels are estimated using Poisson generalized linear models (GLM)40,58. The inputs (i.e., regressors) to the model depend on two events that occur on each trial: onset of fixation (i.e., when the rat inserts its nose into the center port), and the time when the rat leaves the center port and begins to move toward the side port. An impulse (i.e., delta function) at the time of each event is convolved with a linear filter, or kernel, to parametrize the time-varying input related to that event. At each time step, the sum of the inputs is fed through a rectifying nonlinearity (softplus) to specify the neuron’s Poisson firing rate at that time. Three kernels, related to fixation, leftward movement, and rightward movement, are learned by maximizing the marginal likelihood40. b, Example neuron. Two GLM variants were fitted to the same neuron, and for each GLM variant, the observed peri-event time histogram (PETH) is overlaid the cross-validated, model-predicted PETH. The choice-dependence of the PETH of this neuron is well captured by the model variant whose peri-movement kernels start −3.0 s before and 0.5 s after movement onset (left), but less well captured by another variant whose peri-movement kernels time base are limited to −0.5 to 0.5 s (right). c, To identify the optimal start of the movement kernel for each neuron, cross-validated (5-fold) model comparison was performed on seven model variants that vary in the start time of the movement kernels and the number of radial basis functions used to parametrize the kernels. The end time of the movement kernel (0.5 s), and the parametrization of the fixation-related kernel (−1.5 s to 2.0 s and 4 basis functions) are identical for all variants. d, The out-of-sample log-likelihood is highest for the model variant whose peri-movement kernels start at −3.0 s. e, For each neuron, the GLM variant with the highest out-of-sample log-likelihood determines the optimal start of the peri-movement kernels. The mode of the distribution is at −3.0 s. f, The start of peri-movement kernels for most neurons precede the time of the first click. g, The start of peri-movement kernels for most neurons precede the earliest commitment time inferred from MMDDM.
