home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!cs.utexas.edu!sdd.hp.com!caen!sol.ctr.columbia.edu!ucselx!network.ucsd.edu!sdcc12!cs!demers
- From: demers@cs.ucsd.edu (David DeMers)
- Newsgroups: comp.ai.neural-nets
- Subject: Re: Network Inversion
- Message-ID: <37157@sdcc12.ucsd.edu>
- Date: 23 Aug 92 04:27:03 GMT
- References: <BtCqHo.AtH.1@cs.cmu.edu> <37125@sdcc12.ucsd.edu> <al.714491544@jargon>
- Sender: news@sdcc12.ucsd.edu
- Organization: =CSE Dept., U.C. San Diego
- Lines: 51
- Nntp-Posting-Host: beowulf.ucsd.edu
-
- Thanks for the list of references!
-
- In article <al.714491544@jargon> al@jargon.gmd.de (Alexander Linden) writes:
- >demers@cs.ucsd.edu (David DeMers) writes:
-
- >>In article <BtCqHo.AtH.1@cs.cmu.edu> tjochem+@CS.CMU.EDU (Todd Jochem) writes:
- >>>I'm looking for references to network inversion papers. The basic idea
-
- >The method Todd asked for is gradient descent in input
- >space with respect to some error measure at the output.
-
- That's not apparent from his post.
-
- If a network is computing y = f(x), then to my mind inversion
- is finding f^{-1}(y). This is normally one-to-many, and
- if x is in X a compact n-dimensional manifold and y is in Y,
- a m-dimensional manifold, the inverse will be a finite
- set of n-m dimensional manifolds (assuming f is smooth).
-
- If you have an inverse problem, that is, find x such
- that f(x) = y, you can solve it differentially, which seems to be
- what you are calling "inversion". This can be done in several ways,
- either pseudo-inverse methods or using the adjoint operator,
- which is what you get by backpropagating delta_y to the inputs.
-
- This is pretty well known in the control literature.
-
- The differential methods result in a solution on the same
- manifold you started, since they can't drive through a
- critical point (where the Jacobian of the forward map loses rank).
-
- ...
-
- >I did not see that this method calculates the gradient in input space.
-
- I fail to see why a gradient is mandatory...
- Aren't we seeking a solution x to f(x) = y?
-
- >A direct calculation of the input seems not possible because it is
- >often a many to one--mapping. Also the thresholds are not considered in
- >his work...
-
- Sure, it's possible - you just have to select one. It's a
- regularization problem, really.
-
-
- --
- Dave DeMers ddemers@UCSD demers@cs.ucsd.edu
- Computer Science & Engineering C-014 demers%cs@ucsd.bitnet
- UC San Diego ...!ucsd!cs!demers
- La Jolla, CA 92093-0114 (619) 534-0688, or -8187, FAX: (619) 534-7029
-