Tech Report Available

[MJ_CARTE%UNHH.BITNET@vma.CC.CMU.EDU]

***** Please Do Not Distribute to Other Bulletin Boards *****
***** Please Do Not Distribute to Other Bulletin Boards *****

The following technical report may be requested in hardcopy by sending
e-mail to

e_mcglauflin at unhh.bitnet

and specifying UNH Intelligent Structures Group Report ECE.IS.90.03, or
by sending physical mail to

Michael J. Carter
Dept. of Electrical and Computer Engineering
University of New Hampshire
Durham, NH 03824-3591

Abstract follows:

"Slow Learning in CMAC Networks and Implications for Fault Tolerance"
by M.J. Carter, A.J. Nucci, E. An, W.T. Miller, and F.J. Rudolph
Intelligent Structures Group
Dept. of Electrical and Computer Engineering
University of New Hampshire
Durham, NH  03824

     The overlapping structure of receptive fields in the CMAC network
can produce situations in which learning is unusually slow.  It is shown
that sinusoidal functions with spatial frequency near certain critical
frequencies are particularly hard to learn with conventional CMAC networks.
Moreover, the resulting learned weights have significantly greater RMS value
near these critical frequencies, and this poses some concern for network
fault tolerance.  It is then demonstrated that CMAC networks using tapered
receptive fields often exhibit faster learning near the critical frequencies,
and the resulting learned weights can have smaller RMS value than those obtained
using the conventional CMAC.


**** Please Do Not Distribute to Other Bulletin Boards ****

Mike Carter