< Back to Article

A probabilistic, distributed, recursive mechanism for decision-making in the brain

Fig 5

Mapping of rMSPRT computations to the cortico-basal-ganglia-thalamo-cortical loops.

(a) Mapping of the negative logarithm of rMSPRT components from Fig 2. Sensory cortex (e.g. MT) produces fresh evidence for the decision, delivered to sensorimotor cortex in C parallel channels (e.g. MT spike trains). Sensorimotor cortex (e.g. LIP or FEF) computes in parallel the simplified log-likelihoods of all hypotheses given this evidence and adds log-priors—or fed-back log-posteriors after the delay Δ has elapsed. It also adds a hypothesis-independent baseline comprising a simulated constant background activity (e.g. from LIP before stimulus onset) and a time-increasing term from the interaction with the thalamus. The basal ganglia bring the computations of all hypotheses together into new negative log-posteriors (the output of the model basal ganglia; see S3 Fig for details) that are then tested against a threshold. Finally, the thalamus conveys the updated log-posterior from basal ganglia output to be used as a log-prior by sensorimotor cortex. Thalamus’ baseline is given by its diffuse, hypothesis-independent feedback from sensorimotor cortex. (b) Corresponding formal mapping of rMSPRT’s computational components, showing how Eq 9 decomposes. All computations are delayed with respect to the basal ganglia via the integer latencies δpq, from p to q; where p, q ∈ {y, b, u}, y stands for the sensorimotor cortex, b for the basal ganglia, and u for the thalamus. Δ = δyb + δbu + δuy with the requirement Δ ≥ 1.

Fig 5