Article View: triangle.talks
Article #32Triangle NN talk: Klenin
From: Jonathan A. Marshall
Date: Tue, 29 Sep 1992 15:08
Date: Tue, 29 Sep 1992 15:08
32 lines
1751 bytes
1751 bytes
From: <sutton@eos.ncsu.edu>
Date: Tue, 29 Sep 92 09:54:09 EDT
Triangle Area Neural Network Society
PLACE: BNR, Inc. (directions below), Research Triangle Park, NC
DATE: Thursday, October 8, 1992 -- 5:30 PM. Refreshments will be provided.
SPEAKER:
> Prof. MARJORIE KLENIN
> Department of Physics
> North Carolina State University
"EVALUATING GENERALIZATION PERFORMANCE: NOISE SOMETIMES HELPS"
A new method is suggested to estimate the limits of learning imposed on
backpropagation networks by the presence of sampling error. A criterion for
optimal match between network function and data is obtained by requiring
that their consistency not exceed the internal consistency of the training
set; the techniques for doing this are borrowed from thermostatistics. To
illustrate its use, both in identifying the onset of overtraining and in
comparing models of different complexity, several toy problems and small
numerical examples are discussed.
DIRECTIONS: From Raleigh/Durham/Chapel Hill: take I-40 toward Research
Triangle Park and get off at Davis Drive. Turn left (south), and follow
Davis through to the intersection with Highway 54. Continue for another 1/2
mile, and turn left into the BNR entrance on Davis. This is the employee
lot; park there and go in the door at that entrance.
QUESTIONS: Please contact John Sutton, (919)515-5065, sutton@eos.ncsu.edu
Technical Details
Message-ID:
<15725@borg.cs.unc.edu>Path:
arc1-ca.pugleaf.net!sparky!uunet!concert!borg!marshall!marshallShow Full Headers
Complete Article Headers
Path: arc1-ca.pugleaf.net!sparky!uunet!concert!borg!marshall!marshall From: marshall@marshall.cs.unc.edu (Jonathan A. Marshall) Newsgroups: triangle.talks Subject: Triangle NN talk: Klenin Message-ID: <15725@borg.cs.unc.edu> Date: 29 Sep 92 15:08:57 GMT Sender: news@cs.unc.edu Followup-To: triangle.talks Distribution: nc Lines: 33