All-TNNs with six topographic layers of different dimensions and kernel sizes, followed by a fully connected category readout (565 ecoset categories) with softmax. In each layer, units are arranged retinotopically on a 2D sheet. Sets of neighbouring units on the sheet share the same receptive field in the layer below them. A smoothness loss encourages similar weights in neighbouring units on the sheet (as illustrated for one unit by blue arrows). Network training optimizes a composite objective, consisting of a classification cross-entropy loss and the smoothness loss, which is scaled by a tunable factor .
