Fig. 4: A simplified model captures discovered dynamics and diverse neuronal profiles.
From: Transitions in dynamical regime and neural mode during perceptual decisions
a, The velocity vector field of both the discovered dynamics and the DDM line attractor can be partitioned into evidence accumulation (EA) and decision commitment (DC) regimes. b, The MMDDM, a simplified model of the discovered dynamics. As in the behavioural DDM, momentary evidence (u) and noise (η) accumulate over time in the decision variable (z) until z reaches either the left (−B) or right (+B) bound. At this moment, the animal commits to a decision: z becomes fixed and unresponsive to further input. Also at this moment, the encoding weight (w) of each neuron shifts from wEA to wDC, changing how z maps to the predicted Poisson firing rate y through softplus nonlinearity h and baseline b. c, MMDDM captures heterogeneous single-neuron profiles. A ramp PSTH arises when wEA and wDC are equal. d, A decay profile emerges when wDC is zero because, over time, more trials reach the bound where encoding of z vanishes. e, A delay profile results from setting wEA to zero because, early in the trial, it is unlikely to have reached the bound. f, ‘Flip’ is produced by setting wEA and wDC to have opposite signs. g, MMDDM captures heterogeneity in single-neuron temporal profiles. Shading represents 95% bootstrap CI of the mean; the solid line is the model prediction. h, MMDDM has a higher out-of-sample likelihood than a one-dimensional DDM without a neural mode switch. i, MMDDM achieves a higher goodness-of-fit R2 value of the choice-conditioned PSTHs. h,i, P values were computed using two-sided sign tests. j, Model prediction (pred.) and observed psychometric function for one example session. The shaded areas are the 95% bootstrap CI of the mean; the solid line is the model prediction.
