Fig. 1

Ensemble coding in natural scenes. Example of isotropic (uniform) and weighted (anisotropic) averaging in a simple pooling scheme. A Uniform: local features (e.g., the orientations of the trees) are pooled together over the entire visual field, independently of their location and distance from the fovea (the orange fixation spot). In uniform pooling, the activity of many neurons encoding local features (depicted as a set of Gaussian tuning functions) is equally weighted and integrated over large areas of space. Integration (Σ) results in a single estimate: the average orientation of all the trees. B Weighted averaging: integration occurs in anisotropic space, due to changes in the density and size of receptive fields (illustrated by gray circles). In this case, ensemble coding is biased towards features at the center of the scene, because of the higher number and narrower tuning of neurons at the fovea. Hence, much more information is received from the central visual field. C Anisotropies due to spatial biases (here, a bias towards the center and the left-hand side is illustrated by the difference in blurring between the central and left-hand region and the rest of the image). Features inside the region of preferential processing (black tuning functions and lines) are weighted more than features outside (gray tuning functions and lines). Note that the three scenarios can lead to different average estimates (the oriented line resulting from integration)