home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!mcsun!uknet!rsre!breeze!heading
- From: heading@breeze.rsre.mod.uk (Anthony Heading)
- Newsgroups: comp.ai.neural-nets
- Subject: Re: Metrics in Self Organizing Maps
- Message-ID: <BsxKE1.H5n@breeze.rsre.mod.uk>
- Date: 13 Aug 92 16:26:01 GMT
- References: <1992Aug6.145942.4330@vu-vlsi.vill.edu> <BsvLn2.9HM@breeze.rsre.mod.uk>
- Organization: Defence Research Agency
- Lines: 25
-
- Whoops. In an earlier reply, I wrote that for the city-block distance,
- delta w = k sgn(x-w)x. This is indeed wrong, as a couple of my
- colleagues have been kind enough to point out. The correct rule is
- delta w = k sgn(x-w) and the x I included is utterly spurious. Back
- to primary school arithmetic lessons for me, I think...
-
- Raman Rajasekhar notes that this isn't really a terribly good
- update rule, since determining k is tricky for continuous updates.
- I agree. It seems to me that there is a fairly obvious batch
- update method though, which we obtain by requiring
- sum over patterns (sgn(x-w)) = 0. This criterion is satisfied
- when w is the centre of the data in the sense that the number
- of samples which are respectively larger and smaller are equal.
- (For odd numbers of data items w should sit squarely on the
- central one.) Each dimension can be treated independently here.
- I can't see any easy way to find the correct k for a continuous
- scheme at the moment. Any ideas?
-
- Anthony
-
- --
- Anthony Heading (heading@hermes.mod.uk)
- DRA Malvern, UK
- Any opinions expressed are my own and in no way
- represent those of my employer or any other organization.
-