Extended Data Fig. 3 |. COIN model parameters.

Left column: Parameters for illustrating the COIN model (I: purple), model validation (V: brown) and fits to individuals in the spontaneous (S: blue) and evoked (E: green) recovery experiments, to the average of both groups (A: cyan), and individuals in the memory-updating experiment (M: red). Right: scatter plots for all pairs of parameters for the six groups. The overlap of data points suggest parameters are similar across experiments. σ_q: process noise s.d. (Eq. 3); μ_a, σ_a: prior mean and s.d. for context-specific state retention factors (Eq. 10); σ_d: prior s.d. for context-specific state drifts (Eq. 10); α: concentration of local transition probabilities (Eq. 8); ρ: self-transition bias parameter (Eq. 18); σ_m: motor noise s.d. (Eq. 17); α^e: concentration of local cue probabilities (Eq. 9). Parameters used in the figures is as follows. I: Fig. 1 and Extended Data Fig. 1c–e. V: Extended Data Fig. 2a–b. S: Fig. 2c, Extended Data Fig. 6f (column 1) and Extended Data Fig. 2d. E: Fig. 2e, Extended Data Fig. 6f (column 3) and Extended Data Fig. 2e. S & E: Extended Data Fig. 2c. A: Fig. 2b & d, Extended Data Fig. 5 and Extended Data Fig. 9 (bias added for visuomotor rotation experiments: Extended Data Fig. 5a–j,p–s and Extended Data Fig. 9e–l). M: Fig. 3 and Extended Data Fig. 7a–d. S, E & M: (all parameters, but α^e): Fig. 4 and Extended Data Fig. 8. The robustness analyses (Extended Data Fig. 4) used perturbed versions of the same parameters as the corresponding unperturbed simulations. To reduce the number of free parameters in the model, we set the parameters of the hierarchical Dirichlet process that determine the expected effective number of contexts or cues, γ (Eq. 7) and γ^e (Eq. 9), respectively, both to 0.1, the prior mean for context-specific state drifts, μ_d, to zero (Eq. 10), and the standard deviation of the sensory noise, σ_s, to 0.03 when fitting or simulating the model, with the variance of the observation noise (Eqs. 5 and 19) being set to ${\sigma }_{\text{r}}^2={\sigma }_{\text{s}}^2+{\sigma }_{\text{m}}^2$. For visuomotor rotation experiments (Extended Data Fig. 5a–j,p–s and Extended Data Fig. 9e–l), we set the mean of the prior of the bias μ_b to zero (Eq. 20), and its s.d. σ_b to 70⁻¹.