Paired-samples t values contrast classification accuracies of predictable and random tones after training a classifier to capture bottom-up (upper row) and top-down (lower row) prediction-related patterns. Warmer colors reflect higher decoding accuracies when testing on the predictable tones. Positive training times are presented for the bottom-up analysis while negative training times are presented for the top-down analysis. 0 ms testing time reflects the onset of the most probable tone. Interestingly, analysis of bottom-up patterns reveals that feature-specific preactivations are more strongly present in predictable sequences, and even during sleep. At the same time, top-down patterns related to the implicitly learned associations of the predictable transitions are only present in wakefulness. However, more basic predictions, such as prediction violations, reflected by significant post-stimulus patterns in the top-down analysis, are preserved in sleep. Black frames illustrate significant clusters after non-parametric cluster-based permutation analysis. See also Figure S3.
