How Precise Is Predictive Coding for Things We Hear? Mismatch Negativity With Shepard Tones

Predictive coding is a theory that each hierarchical level of the nervous system constructs a model from bottom-up information (such as sensory inputs) and from top-down information (such as expectations and reliability of past information) and then tests new inputs against the model. If the new inputs match the model, then no change in it is required. If not, then extra brain processing is required to update the model. We tested the precision of such a model in the auditory system by using Shepard tones arranged into a discrete Shepard scale-a series of notes, each comprising sine tones of different amplitudes and one octave apart, and with each tone separated from the next by one semitone, yielding a scale that ascends or descends forever. We unpredictably and occasionally replaced an expected tone in the scale by one that was two thirds of a semitone less, one third of a semitone less, one third of a semitone more, or two thirds of a semitone more. We measured the electrical activity of 20 participants' brains with 128 scalp electrodes (electroencephalography, EEG) while the tones were delivered to their ears. We found that event-related potentials (ERPs) from 180-220 ms to these unpredictable tones were more negative the farther they were from the predicted note and more negative for tones that were less than the expected note. We conclude that the predictive model for the kind of regularity in a discrete Shepard scale has a sensitivity to tones less than one third of a semitone and is more sensitive to undershoots than to overshoots.