Fig. 8: Summary of the findings.

The top row is a representation of a set of orientation-selective units (here, columns) processing low-variance inputs, while the bottom row schematizes the processing of high-variance inputs. In the case of low-variance inputs to V1, the underlying sensory distribution is sharp in the orientation space, driving mostly a single orientation-selective unit that processes orientation in a fast feedforward manner. The feature encoded by this activity then stays stable through time (from left to right). For inputs of higher orientation variance, the sensory input is broadly distributed in the orientation space, which drives many dissimilarly tuned units, thus recruiting slow recurrent interactions. The quality of feature encoding progressively increases through time, as recurrent interactions perform computations to represent the most salient oriented feature in the input.