The consequence is a sparse, efficient representation (mostly in

predictor neurons) of predictable input, and a robust, distributed response (mostly in error neurons) to unpredictable input, both coordinated across multiple levels of the processing hierarchy (Figure 1). Within a cortical region, population activity reflects a mixture of responses in the predictor neurons (passing information about predicted inputs down the hierarchy) and the error neurons (passing information about unpredicted inputs up the hierarchy). In principle, predictive coding models need make no assumption about the distribution of these two kinds of neurons within a population; in practice, aggregate population activity is often dominated by error neurons (Friston, 2009, Wacongne et al., 2012, Egner et al., 2010, Keller et al., selleck chemicals 2012 and Meyer and Sauerland, 2009). The result is that the

classic signature of predictive coding, reduced activity to predictable stimuli, is typically observed when averaging across large samples of neurons selleck chemicals llc within a region (Meyer and Olson, 2011, Egner et al., 2010 and de Gardelle et al., 2013). However, (as described in more detail below) signatures of the predictor neurons can also be observed; for example, the predictor neurons would likely show increased response when the input matches their predictions (e.g., de Gardelle et al., 2013). Following work in sensory processing (e.g., Wacongne et al., 2012), in our proposal both error neurons and predictor neurons convey “representational” information, and both are likely tuned to specific stimuli or stimulus features. Predictor neurons, present at each level of the cortical hierarchy, do not code a “complete” representation of the expected stimulus, but only some features or dimensions of the stimulus, at a relevant level of processing. Each set of predictor neurons can explain only those particular features or dimensions of the input, and correspondingly modulates the response in a highly specific subset of error neurons. Error neurons are similarly distributed throughout the cortex and respond to specific stimulus features (Meyer and Olson, 2011 and den Ouden et al., 2012),

rather Tryptophan synthase than, for example, a single “error region” signaling the overall amount of error or degree to which the observed stimulus is unpredicted (e.g., Hayden et al., 2011). Thus, for example, in the early visual cortex, predictor neurons code information about the predicted orientation and contrast at a certain point in the visual field, and error neurons signal mismatches between the observed orientation and contrast and the predicted orientation and contrast. In IT cortex, predictor neurons code information about object category; error neurons signal mismatches in predicted and observed object category (den Ouden et al., 2012 and Peelen and Kastner, 2011). One consequence of this model is that, typically, the effects of predictions are limited to relatively few levels of the processing hierarchy.