Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!ucbvax!janus.berkeley.edu!nadi
From: nadi@janus.berkeley.edu (Fariborz Nadi)
Newsgroups: comp.ai.neural-nets
Subject: Statiscal weights for back-prop nets?
Message-ID: <29109@ucbvax.BERKELEY.EDU>
Date: 8 May 89 23:20:05 GMT
Sender: usenet@ucbvax.BERKELEY.EDU
Reply-To: nadi@janus.berkeley.edu (Fariborz Nadi)
Organization: University of California, Berkeley
Lines: 13


Training algorithm for back-prop networks with statistical weights.

In order to create statistical models by observing input and output vectors
alone, a back-prop network needs to have a statistical component added to it.
This could be done by having statistical weights, thresholds, or nodes.
The reason for such models is to cope with noisy data in the input and output
vectors. Is anyone working on a training algorithm for such networks?
Is there any other way of creating such models without knowing the distribution
of the noise before hand? (remember I am interested in multi-input multi-output
vectors) Any and all feedbacks are appreciated.

nadi@janus.berkeley.edu