(a) The Gaussian gain applied to Layer 1 units caused the measured receptive field (RF) of units in Layer 4 to shift (black ellipse, original; brown ellipse, with gain) toward the locus of attention (black ×). (b) A 2D spatial map demonstrates the effects of Gaussian gain in Layer 4: shift of RF center position (black arrows), shrinking RF size near the attended locus (blue colors) and an expansion of size near the gain boundaries (red colors). (c) map of the output layer before averaging, showing the change in AUC caused by the addition of Gaussian gain. Each pixel’s ΔAUC is computed by projecting the activations at that location for composite grids with target present and absent on the decision axis and then calculating the difference in AUC between a model with and without Gaussian gain. The map demonstrates that units overlapping the borders of the composite grid have the largest change in information content when Gaussian gain is applied. (d,e) Scatter plots demonstrate that each layer magnifies the effect of the gain on RF shift and RF size. The RF shift percentages are the ratio of pixel shift at the peak of the curve relative to the average receptive field size, measured as the full-width at half-maximum. (f) Later layers do not magnify the effective gain (shown for an 11× gain), which stays constant across layers. (g) Gain strength influences the size of RF position shifts, RF size (h), and effective gain (i). (j) Adding an additional non-linear normalizing exponent at the output of each layer allows for much smaller gains to be magnified across layers. Markers in all panels indicate individual sampled units from the model. Lines show the LOESS fit for visualization.