Decoding Predictions and Violations of Object Position and Category in Time-resolved EEG

Classic models of predictive coding propose that sensory systems use information retained from prior experience to predict current sensory input. Any mismatch between predicted and current input (prediction error) is then fed forward up the hierarchy leading to a revision of the prediction. We tested this hypothesis in the domain of object vision using a combination of multivariate pattern analysis and time-resolved electroencephalography. We presented participants with sequences of images that stepped around fixation in a predictable order. On the majority of presentations, the images conformed to a consistent pattern of position order and object category order, however, on a subset of presentations the last image in the sequence violated the established pattern by either violating the predicted category or position of the object. Contrary to classic predictive coding when decoding position and category we found no differences in decoding accuracy between predictable and violation conditions. However, consistent with recent extensions of predictive coding, exploratory analyses showed that a greater proportion of predictions was made to the forthcoming position in the sequence than to either the previous position or the position behind the previous position suggesting that the visual system actively anticipates future input as opposed to just inferring current input.

The Relation between Alpha/Beta Oscillations and the Encoding of Sentence induced Contextual Information

Within the sensory domain, alpha/beta oscillations have been frequently linked to the prediction of upcoming sensory input. Here, we investigated whether oscillations at these frequency bands serve as a neural marker in the context of linguistic input prediction as well. Specifically, we hypothesized that if alpha/beta oscillations do index language prediction, their power should modulate during sentence processing, indicating stronger engagement of underlying neuronal populations involved in the linguistic prediction process. Importantly, the modulation should monotonically relate to the degrees of predictability of incoming words based on past context. Specifically, we expected that the more predictable the last word of a sentence, the stronger the alpha/beta power modulation. To test this, we measured neural responses with magnetoencephalography of healthy individuals (of either sex) during exposure to a set of linguistically matched sentences featuring three distinct levels of sentence context constraint (high, medium and low constraint). We observed fluctuations in alpha/beta power before last word onset, and also modulations in M400 amplitude after last word onset that are known to gradually relate to semantic predictability. In line with previous findings, the M400 amplitude was monotonically related to the degree of context constraint, with a high constraining context resulting in the strongest amplitude decrease. In contrast, alpha/beta power was non-monotonically related to context constraints. The strongest power decrease was observed for intermediate constraints, followed by high and low constraints. While the monotonous M400 amplitude modulation fits within a framework of prediction, the non-monotonous oscillatory results are not easily reconciled with this idea.