This Neural Net example implements a consultation module for a 3D-Kohonen Map. "3D" refers to the cluster space - during training, clusters were allowed to form on a 3D- rather than a 2D- Map.

The network was trained on patterns representing 5x7 bitmaps of the uppercase Alphabet (english). Latteral inhibition was implemented using a "sombrero" form of Kohonen's algorithm. The map used was defined as 3x3x3 (27 nodes) - so, for 26 training patterns, there is one node unassigned in the trainned net. The clustering (node/pattern assignments) formed during training is represented by the letter that appears on each node of the map display.

The objective of this sort of network is to classify an input pattern according to some set of known classifications (patterns). Thus, if the pattern for "A" is presented, the node located at row 1 collumn 3 of layer 0 would be expected to respond (or fire). Likewise, if an altered form of the A-pattern is presented, we would, again, hope that the A-node would respond, letting us know that the pattern is probably an A.

NOTE: Please note that, in this example implementation, the dead node in the map is not allowed to fire.

The active pattern is displayed in the Pattern Viewer/Mangler frame. You may select a training pattern by selecting the appropriate radio button. You can alter the active pattern by clicking on cells of the pattern matrix. You can toggle cells as ON (Black) or OFF (white) by clicking on the cell. Once you are satisfied with the pattern, you can consult the net with the activew pattern by clicking on the Cycle the Pattern button.

The winning node (the node that reacts most strongly to the input pattern), will be highlighted in Red once the pattern has been processed. Also, the Scaled Absolute Error between the Winner's input weights and the pattern values is shown graphically.