Paasch, Georg-Friedrich (2018)
Investigating the Predictive Code of Predictive Coding Theory via Face Identity Expectation in a Learning Task.
Technische Universität Darmstadt
Dissertation, Erstveröffentlichung

Kurzbeschreibung (Abstract)

Human behaviour is based on a reliable recognition of its environment. The predictive coding theory (PCT) provides an explanation of how recognition becomes reliable by integrating prior experience about the environment. The PCT suggests that the brain provides a prediction based on prior experience which is compared with the sensory information. The sensory information that cannot be explained by the prediction causes the prediction to be updated.

An established view is that the brain is organised in a hierarchical manner and the sensory information is transferred bottom-up the hierarchy. Within the hierarchical organisation, the PCT proposes that the prediction is sent top-down to be compared to the sensory information and that the unexplained information, the prediction error (PE), is sent bottom-up to update the prediction (Friston, 2003, 2005; Rao and Ballard, 1999). Accordingly, the information flow circulates: PEs drives Prediction updates, and Predictions drive PE minimization.

The predictive coding concept is compelling, but the neurophysiological implementation still remains unknown. However, Bastos and colleagues (Bastos et al., 2012) suggest that high frequency band activity (>30Hz) reflects PEs and low frequency band activity (<30Hz) reflects Predictions. This variant of the implementation of the PCT is testable with neurophysiological methods like the magnetoencephalography (MEG).

The present thesis investigated the representation of the Prediction within the framework of the PCT with a new learning paradigm using two-tone stimuli. In this paradigm, the participants expected a target face identity (TFID) at a predefined position in a trial to gain reliable visual information of the TFID. Furthermore, the participants were supposed to test their recognition ability with a test stimulus as well at a predefined position in a trial.

First, our understanding is that to solve this task sensory evidences of the TFID must be integrated into a coherent internal model (IM). Second, according to Bubic (Bubic, 2010), we can assume that the IM should become the predictive information when expecting an upcoming stimulus of the TFID. Third, we suppose that the IM should increase in detail through learning and accordingly we assume that this refinement of the IM can be associated with the increasing precision of predictions.

Indeed, an increase in the discrimination performance of the TFID with increasing refinement of the IM could be verified when analysing the behaviour of 41 participants. At the neurophysiological level, this was accompanied by an enhancement of the well-known familiarity ERF component M250 (Olivares et al., 2015; Schweinberger and Neumann, 2016).

Recent proposals of the PCT claim a distinction of the low frequency band activity in alpha and beta frequency band activity (ABA, BBA). It is proposed that the ABA signals the precision of predictions and the BBA signals the update of predictions (Sedley et al., 2016). In our MEG data, we observed a change in the ABA in the expectation interval which had a positive relationship to the increasing discriminability of the participants. Three brain sources of major importance were identified for this effect in the expectation interval and in the ABA by a beamformer source localization approach. These were the occipital face area (OFA), the precuneus (PreC) and the lateral occipital cortex (LOC). These brain areas (and two others) were supposed to constitute a face-identity-predictive (FIP) network. Further, we tested for effects of ABA in the post-stimulus interval and found an increase in ABA in primarily early visual areas and the PreC. Thus, we assign the PreC a key role in the representation of the IM of faces.

It stands to reason that the ABA in the expectation interval and the post-stimulus interval are related to the same IM. Thus, we propose the ABA in the expectation interval to reflect an abstract representation and the ABA in the post-stimulus interval to reflect a more detailed internal model representation. This interpretation is in line with a tentatively proposed implementation of the PE minimization by Kwisthout and colleagues (Kwisthout et al., 2017).

Frage zum Eintrag

Redaktionelle Details anzeigen