Computations that sustain neural feature selectivity across processing stages.

Biological visual systems are celebrated for their ability to reliably and precisely recognize objects. However, the specific neural mechanisms responsible for this capability remain largely elusive. In this study, we investigate neural responses in the visual areas V1, V2, and V4 of the brain to natural stimuli using a framework that includes quadratic computations in order to capture local recurrent interactions, both excitatory and suppressive. We find that these quadratic computations and specific coordination between their elements strongly increase both the predictive power of the model and the neural selectivity to natural stimuli. Particularly important were (i) coordination between excitatory and suppressive features to represent mutually exclusive hypotheses regarding incoming stimuli, such as orthogonal orientations or opposing motion directions in area V4, (ii) balance in the contribution of excitatory and suppressive components and its maintenance at similar levels across stages of processing, and (iii) refinement of feature selectivity between stages, with earlier stages representing broader category of inputs. Overall, this work describes how the brain could use multiple nonlinear mechanisms to increase selectivity of neural responses to natural stimuli.