Evidence for predictive coding in early visual cortex in context of self-induced visual motion

The model of “predictive coding” suggests that feedback from a higher- to a lower-order visual area carries predictions of lower-level neural activities, whereas the feedforward connections carry the residual errors between the predictions and the actual lower-level activities (Rao and Ballard, 1999). We tested this theory in context of processing of planar motion in early (foveal) visual cortex. In a 2x2 factorial design, human subjects either fixated (eyes still) or carried out smooth pursuit on a display containing a planar random dot-field that was either stationary or moving coherently in-plane. This led to four conditions: (a) fixation on static dot-field, (b) fixation on moving dot-field, (c) pursuit on static dot-field, (d) pursuit of moving dot-field (pursuit was locked to the dot-motion). Neural activity was measured using fMRI at 3T. If early visual cortex coded for retinal motion, (b) and (c) would be expected to activate early visual cortex equally, and more than (a) and (d). In contrast, predictive coding would result in different responses. In addition to the above, early visual cortex would also code the error signal for mismatches between retinal motion input and the prediction for retinal motion, based on e.g. pursuit-related efferent copies. Such mismatches between prediction and input would occur in (b) (retinal motion without prediction of it) and in (d) (absence of retinal motion despite prediction of it). Note that these mismatches are equivalent to the presence of objective motion in the display. Thus, predictive coding would lead to highest responses in (b) (error + input), medium responses in (c) (input only) and (d) (error only), and lowest response in (a). We found (across the whole brain) the only activity satisfying these criteria in the occipital poles. The occipital poles contain the foveal confluence of early visual areas V1-V3, and are thus the key candidate for the above hypothesis. Their responses matched the hypothesized pattern precisely. In contrast, activity in motion responsive areas such as V5/MT+ and parietal regions was mainly driven by eye-movements and by retinal motion. Offline eye-tracking revealed that our results cannot be explained by differential fixation accuracies across conditions. It remains to be elucidated whether predictive coding actually accounts for the results, or whether direct feedback of objective motion signals from higher-level areas sums up with retinal input to the response observed in the occipital pole. Nevertheless, our results let us conclude that activity in the foveal representation of the early visual cortex fully match the predictions of Rao and Ballard (1999) for predictive coding