Figure 4.

The BNN Is Optimized for the Translation of Image Features that Arises from the Geometry of Binocular Viewing

(A) Computing the optimal stimulus for a complex unit. Starting with random noise inputs, the algorithm computed the gradient of complex unit activity with respect to the input images. It iteratively adjusted the inputs to maximize the complex unit’s activity.

(B) Snapshots of three iterations during optimization: a consistent on-off pattern emerges in the left and right eyes, horizontally translated to match the preferred disparity of the unit.

(C) This pattern remains when “lesioning” the BNN of 25% of the simple units that use position encoding.

(D) Removing highly weighted hybrid units leads to input images that are unrealistic.