Fig. 2: Unsupervised discovery shows transitions in dynamical regime and neural mode underlying the shift from evidence accumulation to decision commitment.

a, Decision-relevant dynamics are inferred using FINDR²⁰. b, FINDR learns the decision variable z that best captures neural spiking activity. Each neuronal spike count at a given time step is modelled as a Poisson random variable with the rate given by an affine transformation of z at that time step, followed by the softplus nonlinearity. The grey box indicates the decision variable z at an example time step, and the yellow box indicates the spike counts at that time step. A time-varying baseline is learnt for each neuron to capture the decision-irrelevant component of its activity. c–h, Vector field inferred from 96 simultaneously recorded choice-selective neurons in the dmFC and the mPFC from a representative session. Only the portion of the state space visited by at least 50 of 5,000 simulated 1-s trajectories (sample zone) is shown. c, Autonomous dynamics. d, Speed of autonomous dynamics. e, Input dynamics for left and right clicks. If u = [1;0] indicates a left click input, F(z, [1;0]) − F(z, 0) is the input dynamics given a left click. However, the average left input dynamics depend on the frequency of left clicks, given by p(u = [1;0]|z). Therefore, we compute the average left input dynamics F(z, left) − F(z, 0) as p(u = [1;0]|z)(F(z, [1;0]) − F(z, 0)). We compute the average right input dynamics similarly, with u = [0;1]. f, Speed of input dynamics. g, Difference in speed between autonomous and input dynamics. h, Initially, z is strongly driven by inputs, and its trajectories develop along the evidence accumulation axis aligned with the direction of input dynamics. At a later time, the trajectories become largely insensitive to the inputs and are instead driven by autonomous dynamics to evolve along the decision commitment axis aligned with the direction of autonomous dynamics.