Abstract level description

Understanding the brain as a prediction machine offers a compelling framework for perception, action, and cognition. Irrespective of the neuronal implementation, the framework ascribes a function to internal models and neuronal processes to best prepare for the anticipated future. At an abstract level, the predictive coding framework also draws attention to two blind spots in neuroscience: (1) internal cortical communication (i.e., maintaining internal models) and (2) the brain processes prior to stimulation onset (i.e., predictive processing).

A starting point to explore internal communication is by investigating cortical feedback (Van Essen Reference Van Essen2005; Muckli & Petro Reference Muckli and Petro2013). Conventional paradigms struggle, however, to isolate cortical feedback during sensory processing (which includes both feedforward and feedback information). We have demonstrated such separation by blocking feedforward stimulation using visual occlusion and reading out rich information content (multivariate patterns) from within non-stimulated regions of the retinotopic cortex (which receive cortical feedback activation; Muckli & Petro Reference Muckli and Petro2013; Smith & Muckli Reference Smith and Muckli2010). By decoding cortical feedback, we begin to shed light on internal processing. With regard to investigating brain processes prior to stimulation onset, we have shown that motion predictions are carried over to new retinal positions after saccadic eye-movements (Vetter et al. Reference Vetter, Edwards and Muckli2012), which confirms that saccadic updating incorporates predictions generated during pre-saccadic perception. This is an important proof of concept of predictive coding in saccadic viewing conditions. Moreover, Hesselmann et al. (Reference Hesselmann, Kell and Kleinschmidt2010), have shown that variations in baseline activity influence subsequent perception, and a causal role of V5 in generating predictions sent to V1 can be demonstrated using transcranial magnetic stimulation (TMS). Pilot data show that TMS interferes with predictive codes during the baseline prior to stimulation onset (Vetter et al., under revision). If the brain would be seen as a “representation machine” instead of a “prediction machine,” one would not look for predictive brain processing before stimulus onset and important information about cortico-cortical communication would remain concealed. Motivating the search for predictive signals in the system is therefore another important contribution of the conceptual framework.

Concrete level description

On the concrete conceptual level, hierarchical cortical prediction provides a scaffold on which we can constrain variants of predictive coding models. Predictions are proposed to explain away the incoming signal or filter away the unexpected noise (Grossberg Reference Grossberg2013). Rao and Ballard (Reference Rao and Ballard1999) proposed a model in which forward connections convey prediction errors only, and internal models are updated on the basis of the prediction error (Rao & Ballard Reference Rao and Ballard1999). Grossberg on the other hand proposes Adaptive Resonance Theory (ART) models that update internal models based on recognition error. It remains an empirical question which combination of these models suffices to explain the rich and diverse cortical response properties. A recent brain imaging study shows that under conditions of face repetition, some voxels show repetition suppression consistent with the concept that the prediction error is reduced with every repetition of the identical image, while others (30%) show repetition enhancement (De Gardelle et al. Reference de Gardelle, Waszczuk, Egner and Summerfield2012). Repetition enhancement in a subpopulation of fusiform face area (FFA) voxels could reinforce the internal model of the face identity and be used to stabilize the prediction. The claim that the brain is a prediction machine might be true regardless of the precise implementation of predictive coding mechanism. Internal models might update on error, stabilize on confirmation or scrutinize on attention (Hohwy Reference Hohwy2012). A recent brain imaging study investigated whether expectation induced signal suppression coincides with sharpening of the underlying neuronal code (Kok et al. Reference Kok, Jehee and de Lange2012). Consistent with the predictive coding framework, auditory-cued stimuli led to reduced V1 fMRI activity. Although the overall activity was reduced, the activation profile was more distinct, “sharpened,” for the expected conditions as measured using multivariate decoding analysis. The study concludes that expectation helps to explain away the signal while attention amplifies the remaining prediction error (Hohwy Reference Hohwy2012; Spratling Reference Spratling2008b).

Another concrete level aspect of predictive coding relates to the question of spatial precision. Are the back-projected predictions at the precision level of the “sending” brain area (i.e., coarse), or at the precision level of the “receiving” brain area (i.e., spatially precise)? We have evidence in favor of both; V5 feedback signals spread out to a large region in primary visual cortex (de-Wit et al. Reference de-Wit, Kubilius, Wagemans and Op de Beeck2012; Muckli et al. Reference Muckli, Kohler, Kriegeskorte and Singer2005) but spatio-temporal predictions in V1 which have been relayed by V5 can also be spatially precise (Alink et al. Reference Alink, Schwiedrzik, Kohler, Singer and Muckli2010). The optimal way to account for this discrepancy is by assuming an architecture that combines coarse feedback with the lateral spread of feedforward signals (Erlhagen Reference Erlhagen2003). If this principle holds true, it helps to explain why the architecture of cortical feedback as described by Angelucci et al. (Reference Angelucci, Levitt, Walton, Hupe, Bullier and Lund2002) contributes to precise predictions even though it is divergent.

The examples above show that on an abstract level important new research is motivated by the hierarchical predictive coding framework and on a concrete conceptual level, the many interactions of cortical feedback of predictions, processing of prediction errors, and different accounts of feedforward connections (some stabilizing the internal model, others explaining away signal discrepancies) await further empirical scrutiny. However, the developing narrative of predictive coding becomes increasingly compelling with attention from sophisticated human neuroimaging and animal neurophysiological studies (Muckli & Petro Reference Muckli and Petro2013). Not only is extending our knowledge of cortical feedback and its encapsulated predictions essential for understanding cortical function, but important opportunities will arise to investigate deviations of predictive coding in aging and neuropsychiatric diseases such as schizophrenia (Sanders et al. Reference Sanders, Muckli, de Millas, Lautenschlager, Heinz, Kathmann and Sterzer2012).