Figure 1: A schematic representation for the QC model.

Each input image is split into overlapping patches. Stimuli from within each patch are passed through a set of linear filters, followed by quadratic nonlinearity (with a linear term). The squared filter outputs are added together with different weights. The weights are positive for excitatory subunits (blue Gabor outlines) and negative for suppressive subunits (red Gabor outlines). Applying logistic transformation yields the response of each QLS. The QLS parameters are the same for all patches. A weighted sum of the QLS outputs across time and spatial patches describes temporal dynamics and graded position invariance, respectively. A rectifying nonlinearity applied to the weighted sum of the QLS outputs yields the predicted firing rate. QLS, quadratic logistic subunit.