home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!brunix!cs.brown.edu!mpp
- From: mpp@cns.brown.edu (Michael P. Perrone)
- Subject: Re: Reducing Training time vs Generalisation
- Message-ID: <1992Aug18.231650.27663@cs.brown.edu>
- Keywords: back propagation, training, generalisation
- Sender: mpp@cs.brown.edu (Michael P. Perrone)
- Organization: Center for Neural Science, Brown University
- References: <arms.714091659@spedden> <36944@sdcc12.ucsd.edu> <arms.714146123@spedden> <36967@sdcc12.ucsd.edu>
- Date: Tue, 18 Aug 1992 23:16:50 GMT
- Lines: 9
-
- The example given of a "wild" solution to a backprop problem
- ( f(x) = 40 [ 1/( 1 + e^40*(x - 1/4)) + 1/( 1 + e^-40*(x - 3/4)) -1 ] )
- is certainly a valid solution. But whether gradient descent from an
- initially "well-behaved" f(x) (e.g. one with suitable bounded derivatives)
- would fall into the "wild" local minima is not clear.
-
- This example is more on the lines of an existence proof than a
- constructive proof: Wild minima can exist but is gradient descent
- likely to converge to them?
-