Talk:Kohonen Map

Some thought on the Kohonen Map. From my recent char with Chase, where he asked me what is the difference between this and k-means clustering. Basically this is more statistical but the k-means is more logged. So he said that this *should* work better against statistical movement (or something like this).

Now I wonder, basically in VCS we set segmentation. How about initialize each node's input with every possible groups of segmentation, then let it adjust dynamically for the input situation. (That is, we simulate each stats buffer in the VCS system with an adjustable node). Any thoughts? --Nat Pavasant 14:59, 10 November 2009 (UTC)

Actually I asked if he was asking me the difference, (he just happened to answer) and note that k-means is the forbearer of DC. My point is that one is more statistical where as the other is more logged. I went on to point out similarities between this and what is currently used as VCS (also namely the rolling as well). Using a non random initialization is very possible and may indeed lead to better outcomes, but if you do I suggest setting the internal time to a higher number before you start training it (atleast 10 or 20), so that it doesn't erase your initialization almost immediately. This network learns linear targeting pretty quickly. --Chase 00:30, 11 November 2009 (UTC)