Fig 1.
(A) Participants performed location and latency discrimination of visual targets. An auditory cue consisted of an ascending (orange) or a descending (blue) pair of tones (cue composition), presented at a high or low pitch. The pitch of the cue could predict the location (left versus right) of the visual target, while the composition of the cue could predict the latency (early versus late) of the target, followed by a speeded discrimination response. Participants performed location or latency discrimination in separate blocks. (B) Cue validity varied unbeknownst to the participants over the course of the experiment. Spatial (blue) and temporal (red) validity levels were uncorrelated and changed implicitly. Alternating tasks (black) were prompted by explicit instructions. (C) Predictability interacted with task relevance in both tasks, improving accuracy when the predictions were relevant. The main effect of relevance reflected the differences in accuracy between tasks. N = 17; error bars: SEM; post-hoc t tests * p < 0.05; ~ p < 0.1. Data pertaining to this figure are available on Figshare https://figshare.com/s/2d2755bfdeea1cbb415f. ISI, inter-stimulus interval; ITI, inter-trial interval; n.s., not significant; RT, reaction time.
Fig 2.
(A) The HGF comprises an observer part, describing the beliefs inferred at 3 levels (low: predictions about target location/latency; middle: cue-target validity level; high: volatility of cue validity), and the response part, linking these beliefs to predicted responses. The full model assumes all 3 levels and a weighted influence of relevant (saturated blue/red) and irrelevant (unsaturated) predictions on participants’ responses. Grey: model states; orange: model parameters. (B) Three alternative observer models (HGF3, HGF2, RW) and 2 alternative response models (task-general: weighted influence of relevant and irrelevant predictions; task-specific: exclusive influence of relevant predictions) were subject to Bayesian model selection. Plot shows log-model evidence relative to the weakest model and indicates task-specific HGF2 as winning. (C) HGF-derived trial-by-trial time-series (representative participant) of predictions about target location/latency (; upper panels) and cue validity (
; middle panels) and PEs about target location/latency (|ε2|; lower panels). (D) Mean correlations between HGF regressors. (E) Correlation between the prior variance of validity level updates and mean accuracy across participants. Data pertaining this figure are available on Figshare https://figshare.com/s/2d2755bfdeea1cbb415f. HGF, Hierarchical Gaussian Filter; HGF2, 2-level HGF; HGF3, 3-level HGF; PE, prediction error; RW, Rescorla-Wagner.
Fig 3.
Neural correlates of predictions and PEs.
(A) Source reconstruction. Auditory cortex (slices centred at MNI 50, −22, 18) and MTG (MNI 62, −40, −8) were identified as main sources of cue-evoked activity. TPJ (MNI 52, −48, 30) differentiated between cue- and target-induced responses. V1 (MNI −2, −100, 2) was the main source of target-evoked activity. (B) Cue-induced prediction correlates are modulated by task relevance. Plots show TF maps of (far left) cue-induced activity independent of any modulation by prediction type or relevance; (mid left) F-contrast across 4 conditions (spatial relevant; spatial irrelevant; temporal relevant; temporal irrelevant), indicating clusters of activity showing significant differences between the conditions; (mid right) T-statistic map of the main effect of relevance, indicating significant differences between relevant versus irrelevant predictions; (far right) parameter estimates per condition for the significant cluster (error bars: SEM). Dashed line marks cue onset. Outlines show F-contrast clusters significant at pFWE < 0.05. (C) Target-induced PE correlates are modulated by task relevance. Plots show TF maps of (far left) target-induced activity independent of any modulation by PE type or relevance; (mid right) T-statistic map of the main effect of relevance, indicating significant differences between relevant versus irrelevant PEs. Dashed line marks target onset. Cluster outlines, mid-left and far-right panels as in (B). Data pertaining this figure are available on Figshare https://figshare.com/s/2d2755bfdeea1cbb415f. A1, primary auditory cortex; MNI, Montreal Neurological Institute; MTG, middle temporal gyrus; PE, prediction error; TF, time-frequency; TPJ, temporoparietal junction; V1, calcarine cortex.
Table 1.
Source reconstruction results.
Table 2.
Effects of trial-by-trial predictions and PEs on TF responses.
Fig 4.
(A) Frequency-by-frequency maps of modulatory effects of contextual relevance on prediction processing. Effects were modelled in a network of 4 interconnected areas, corresponding to the 4 regions in which significant effects of relevance on prediction-related responses were identified (cf. Fig 3B). (B) The corresponding maps of modulatory effects of contextual relevance on PE processing, modelled in a network of 2 areas in which significant effects were identified (cf. Fig 3C). (C) Principal frequency modes estimated for prediction-related responses across the modelled areas. (D) Significant modulatory parameters corresponding to the effects of contextual relevance on prediction-related responses. Each bar represents a significant modulation (by contextual relevance) of the influence of a particular frequency mode in 1 region on another frequency mode in another region. (E) Modulatory spectra of the relevance-related effects of A1 activity (left panel, corresponding to Mode 1) on prediction-induced activity in all regions (right panel, corresponding to an average across frequency modes weighted by the respective modulatory parameters). (F-H) Same as (C-E) but for PE processing (H, left panel: an average of Modes 2 and 3). Data pertaining this figure are available on Figshare https://figshare.com/s/2d2755bfdeea1cbb415f. A1, primary auditory cortex; DCM, dynamic causal modelling; MTG, middle temporal gyrus; PE, prediction error; TPJ, temporoparietal junction; V1, calcarine cortex.
Table 3.
Optimised parameters of the winning HGF model.
Fig 5.
Convolution modelling for TF responses.
TF data from the entire experiment (without epoching) were modelled using a GLM approach, with the design matrix specifying event and nuisance regressors (columns, left to right: cue onset and its 5 modulation regressors; target onset and its 5 modulation regressors; response onset and its 3 modulation regressors; 5 EOG and pupil size nuisance regressors; 6 motion regressors; see main text for details). Because each regressor was modelled as a Fourier time series (inset below), the resulting Fourier coefficients (here depicted for the third out of k columns and corresponding to parameter estimates for m basis functions and f frequencies) constitute a deconvolved TF response to each event type and/or parametric regressor. Data pertaining this figure are available on Figshare https://figshare.com/s/2d2755bfdeea1cbb415f. GLM, general linear model; EOG, electrooculography; freq, frequency; TF, time-frequency.