home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!pmafire!news.dell.com!swrinde!zaphod.mps.ohio-state.edu!saimiri.primate.wisc.edu!ames!agate!doc.ic.ac.uk!warwick!kinguni2!ceres!pd_s001
- From: pd_s001@ceres (J Karwatzki)
- Newsgroups: comp.ai.neural-nets
- Subject: Re: Meaning of deltas in hidden units
- Date: 17 Dec 1992 16:30:14 GMT
- Organization: Kingston University
- Lines: 57
- Message-ID: <1gq9umINN21q@mercury.kingston.ac.uk>
- References: <1992Dec15.162420.22854@Informatik.TU-Muenchen.DE>
- NNTP-Posting-Host: ceres.kingston.ac.uk
- X-Newsreader: Tin 1.1 PL4
-
- baginski@Informatik.TU-Muenchen.DE (Boris Baginski) writes:
- :
- : Everbody knows the meaning of the delta values in the backprop algorithm
- : for output units (difference between target and activation) and the use
- : of the deltas to adjust the internal weights.
- :
- : But the question is: what do those values express for the hidden units and
- : for the input units (can be calculated easily, are not used) ?
- : We explored that these deltas take on values up to 50 (+/-) or so, but we only
- : use inputs and targets and activation states in the range of 0-1.
- : So it`s not simply possible to say these values express a realistic target minus
- : activation for the hidden/input units.
- :
- : thank you, regards,
- : BORIS BAGINSKI
- : adress: baginski@informatik.tu-muenchen.de
-
- Since your targets are set to +/- 1 then you must be using some squashing
- function (sigmoid, tanh etc) on the output activation level (otherwise
- they would take on levels of +/-50 or so). If you want the hidden deltas to be
- in the range +/-1 then use the same squashing function. Whatever you do
- the deltas indicate the relative error of the hidden units backpropagated from
- the output units. Interestingly you can use this error to update the hidden
- activation levels (ignoring activations passed down from higher levels) so
- that you obtain hidden patterns best suited to obtaining the required
- target patterns (for any specific initial set of interconnecting weights).
- This could be useful to find out how far you are from final convergence in
- a multilayer network.
- --
- John Karwatzki
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-