The encoding of stochastic regularities is facilitated by action-effect predictions

Our brains continuously build and update predictive models of the world, sources of prediction being drawn for example from sensory regularities and/or our own actions. Yet, recent results in the auditory system indicate that stochastic regularities may not be easily encoded when a rare medium pitch deviant is presented between frequent high and low pitch standard sounds in random order, as reflected in the lack of sensory prediction error event-related potentials [i.e., mismatch negativity (MMN)]. We wanted to test the implication of the predictive coding theory that predictions based on higher-order generative models—here, based on action intention, are fed top-down in the hierarchy to sensory levels. Participants produced random sequences of high and low pitch sounds by button presses in two conditions: In a “specific” condition, one button produced high and the other low pitch sounds; in an “unspecific” condition, both buttons randomly produced high or low-pitch sounds. Rare medium pitch deviants elicited larger MMN and N2 responses in the “specific” compared to the “unspecific” condition, despite equal sound probabilities. These results thus demonstrate that action-effect predictions can boost stochastic regularity-based predictions and engage higher-order deviance detection processes, extending previous notions on the role of action predictions at sensory levels.


The passive listening task: Method
This part of the experiment was always run after the active task. Participants watched "Home", a documentary that was silently played, while listening to sequences of tones in two separate conditions. In the condition with deterministic regularities (hereafter, DREG) the 900-Hz and 1100-Hz standard tones were presented in alternation while rarely, one of the two standard tones was replaced by the 1000-Hz deviant. The ratio of standard-to-deviant tones was 90% -10%, with 45% for each of the standard tones. In the condition with stochastic regularities (hereafter, SREG), the order of the two standard and deviant tones was random, while the overall high (90%) or low (10%) probabilities of the standard and deviant tones were the same as in DREG. The SREG sequences represented a passive replay of the tones produced in the active task; essentially, every participant listened to the tones that they previously generated in the SPEC condition. The timing in the SPEC condition further determined the stimulus-onset-asynchrony (SOA) in both the SREG and DREG conditions. That is, the tones were presented at the same pace of about one second that every participant had previously generated. Note that trials containing timing errors in the active task were not included in calculating the SOAs, thus a tone was always presented after a minimum of 500 ms and a maximum of 1500 ms following the preceding tone.
Similarly to the active task, the passive listening task consisted of 20 experimental blocks, 10 for every condition, the duration of one block was about 1.5 min, and participants could take self-paced breaks in between. Experimental blocks again consisted of 90 standard tones and 10 deviant tones, 900 and 100 trials being thus collected in every condition for the standard and deviant tones, respectively. The condition order was implemented as follows.
Out of the seven participants who started with SPEC in the active task, four started with DREG, and three started with SREG the in the passive task. Regarding the seven participants who started with UNSPEC in the active task, two started with SREG, and five with DREG in the passive task. In total, nine participants started with DREG and five with SREG, in the passive task. We have initially planned to collect a larger dataset involving full counterbalancing of the four conditions' order across both tasks; however, this was not possible due to the start of the COVID-19 pandemic.
Regarding the EEG data recording and preprocessing, the same steps as for the active task were applied (see main manuscript body). In order to determine the ERP components of interest, a temporal Principal Component Analysis (PCA) was performed on the grandaverage data corresponding to the standard and deviant tones in the DREG and SREG conditions, with the same parameters as in the active task (see main manuscript body for details).

The passive listening task: Results and Conclusion
The grand-average ERPs along with the PCA results are displayed in Supplementary  Supplementary Table S1 presents a summary of the statistical results, which we report next. Figure S1. ERP PCA results. Grand-average ERPs (left) display the standard, deviant and difference waves for the deterministic (a) and stochastic (b) regularities, for an average of frontocentral Fz, FC1, FC2, Cz, CP1, and CP2 electrodes. Following the PCA analysis, 20 principal components explaining more than 95% of the epoch variability were retained, the sum of these components or the so-called reconstruction waves (middle) being displayed again for the standard, deviant, and difference wave in both conditions, for the same average of electrodes as before. Note that the reconstruction waves correspond well to the grand-average ERPs indicating the PCA solution accurately represents the original data. The 20 retained components are presented individually (right); out of these, three components presumably representing N1, MMN, and N2 responses were further analysed.  Supplementary Table 1). Similarly, all Bayesian t-tests corresponding to the main effects and interaction brought moderate to weak support for the null hypothesis (see Supplementary Table 1). Therefore, no N1 effects (nor relevant condition differences) were observed in these data.

MMN:
The frequentist repeated-measures ANOVA lead to a significant main effect of  Table 1). Thus, although the N2 effect seems to be larger in the DREG condition (see Supplementary Figure 2b), the statistical evidence does not support any reliable N2 effects.
To sum up, we found similar MMN responses following the violation of stochastic and deterministic regularities. Note that these effects were clearly post-N1, which was detected as a separate component for which no reliable modulation was found; thus, the observed MMN responses do not represent neural adaptation, but presumably "true" prediction-related processes 1 . Additionally, a non-significant/weak difference between the two conditions at the level of the N2 component indicates that deterministic regularities might still be somewhat easier to recognize. This would be congruent with earlier results from the original study 2 indicating a difference between the two regularity types, especially if we consider that the MMN window in the original study was close to the N2 time range (where the ERP components were not derived via temporal PCA). The non-significant/weak (by contrast to stronger) N2 difference could be further explained by the fact that in this study, in the deterministic condition, a small (unexplained) positivity for the deviants just preceding the MMN and N2 (see Supplementary Figure S2) may have artificially reduced the deterministic mismatch effects and consequently the difference between the two regularity types. In conclusion, it remains for future studies to clarify the differences between stochastic and deterministic regularity encoding. Corresponding Bayesian pairwise comparisons tested the magnitude (or the lack) of the evidence regarding the frequentist main effects and interactions (similarly to the main manuscript analyses). Note that these complementary analyses insure optimal correspondence between the Bayesian and frequentist results, while allowing evaluating support provided by the data for the null hypothesis as well. Significant frequentist effects and BF10 supporting H1 are highlighted in bold.

Analyses of mastoid data
In order to distinguish the MMN from the N2 component, we analysed the mastoid data to check whether the MMN, but not the N2, inverts polarity at these sites 3  To conclude, a polarity inversion at the mastoid sites has been observed for MMN component, but not for N2 component.