home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!synaptx!platt
- From: platt@synaptx.Synaptics.Com (John Platt)
- Newsgroups: comp.ai.neural-nets
- Subject: Interpolation and neural networks
- Keywords: interpolation partition-of-unity
- Message-ID: <13163@synaptx.Synaptics.Com>
- Date: 13 Aug 92 21:58:22 GMT
- References: <arms.713640420@spedden>
- Reply-To: platt@synaptics.com (John Platt)
- Organization: Synaptics Inc., San Jose, CA
- Lines: 54
-
- In an article arms@cs.UAlberta.CA (Bill Armstrong) writes:
- >AN IMPORTANT QUESTION: Can anyone think of any other straightforward
- >ways to get bounds on values *between* test points? The application
- >of NNs in safety-critical areas is at stake.
-
- Yes.. It's quite straightforward and has already been suggested by
- Moody & Darken and separately by Kurt Stockbro for neural networks.
- It's a trick from approximation theory called "partition of unity" or
- sometimes a "generalized Shepard's method".
-
- Normally, when you have a radial basis function network, the network
- output is a linear superposition of radial basis functions:
-
- O_i = sum_j w_ij f_j(input)
-
- Moody & Darken suggest averaging the outputs, not summing them:
-
- sum_j w_ij f_j(input)
- O_i = ---------------------
- sum_j f_j(input)
-
- That way, the output is always in the convex hull of the w_ij for any
- basis function f_j that is non-zero. If the f_j have local support, then
- you can bound the output vector in a region quite easily.
-
- Notice that the meaning of the w_ij has changed using a partition of
- unity. In the superposition case, the w_ij are small vectors that accumulate
- towards the correct answer. In the averaging case, the w_ij are stored
- memories of the O_i and the f_j control how strongly that memory is
- recalled.
-
- Personally, I have found that averaging makes smoother functions than
- superposition.
-
- References:
-
- Moody, J., Darken, C., "Fast Learning in Networks of Locally-Tuned
- Processing Units," Neural Computation, vol 1, no. 2, pp. 281-294.
- (The relevant equation is on page 283), (1989).
-
- Stockbro, K., "Predicting Chaos with Weighted Maps", M. S. Thesis
- Niels Bohr Institute, (1991)
-
- Other References back to Approximation Theory:
-
- Franke, R., "Scattered Data Interpolation: Test of Some Methods,"
- Mathematics of Computation, vol. 38, no. 157, pp. 181-200, (1982).
-
- Shepard, D., "A Two-Dimensional Interpolation Function for Irregularly
- Spaced Data," Proc. 23rd Nat. Conf. ACM, pp. 517-523, (1968).
-
-
-
-
-