The Kohonen Map is also named Self-Organizing Map or SOM. The SOM Color Demo here gets input from three color channels: red, green, and blue. The strength in each channel various from 0 (black) to 1. Each of the neurons in the map is colored by the strength of its weigths. So, if a neuron has input weights (0.5, 0.5, 0.5) it will be shown as mid-tone grey, but if it has (0,1,0), it will be colored bright green.
At each iteration or cycle, a new color sample is randomly generated and presented to the map. The sample and its values are shown at the bottom (use F11 to set a PC browser fullscreen). The best fitting node (with weights most resembling the color sample) 'wins' and its weights are shifted a little bit (further) towards the color sample. Also all nodes around it, within the Radius (in neurons) are similarly shifted toward the sample, making respond a little more to that particular color sample. The size of this shift can be set with learning parameter Rate.
You can step through the learning process, one cycle at a time, by pressing . Or you can press , which runs through the number of cycles, n, set in the Cycles input field. One Cycle here means: Present one randomly generated color pattern, find the winning node, and update the weights to that node and to its neighbors.
Right-clicking a neuron lesions it, turning it bright white. It no longer learns or participates in the cycles. This creates a 'wall' in the map; you can use this to create isolated areas of different shapes.
11.a Let the 10x10 map learn with different settings of radius (1, 3, 5) and rate (0.01, .1, .2). Reset the weights in between runs with (or reload the page). Describe your observations for each of the nine combinations. Comment in particular on the effect of the settings on speed of convergence, stability of the map, and formation of suboptimal representations (e.g., several patches of the same colors). Depending on the learning rate, you may have to set the Cycle parameter to a high value such as 100 or 500 or higher.
11.b Select one setting from 11.a that leads to a stable map. Cycle until you have a converged, stable map. Then, disable one (or two) color channels by unchecking it at the very bottom of the simulator. Now, new samples will no longer contain certain color values (e.g., no colors with 'red'). This correspond to 'amputating' a sensory area or type (in this case: Loss of a type of cone in the retina). Describe your observations (and report the settings used). What happens when you reactivate the lesioned channel? Ponder how this may be a model for how the brain reacts to say the partial amputation of say a finger, which eventually regains its functionality.
11.c Select a small radius (e.g., 1) and a low learning rate (e.g., 0.05). Set Cycles to a high number (e.g., 500). Create a horizontal divide in the map by right-clicking the neurons in the middle from left to right. Give this 'wall' a thickness of two rows. Now you have an upper map and a lower map, which are not in contact. Verify this by learning a few thousand cycles until you see a convergence. Now create a hole in the 'wall' with a width of three (3x2) by left clicking on the lesioned neurons, which will then turn from white to black (and which will then start taking part in the learning process again). Run for another few thousand cycles. What do you observe? Now, turn off one of the color channels and run for another few thousand cycles. What do you observe now? Can you think of what caused the difference?
11.d Think of a interesting variation or application of one of the simulations above or make up your own one. Briefly describe. E.g., you could play with different map sizes, different types of lesion patterns in the map, or different lesion patterns in color channels.