Tech Reports Available

[Eduardo Sontag]

The following report is now available:

          "On the recognition capabilities of feedforward nets"
         by Eduardo D. Sontag, SYCON Center, Rutgers University.

ABSTRACT: In this note we deal with the recognition capabilities of various
feedforward neural net architectures, analyzing the effect of direct input to
output connections and comparing Heaviside (threshold) with sigmoidal response
units.  The results state, roughly, that allowing direct connections or
allowing sigmoidal responses doubles the recognition power of the standard
architecture (no connections, Heaviside responses) which is often assumed in
theoretical studies.  Recognition power is expressed in terms of various
measures, including worst-case and VC-dimension, though in the latter case,
only results for subsets of the plane are proved (the general case is still
open).  There is also some discussion of Boolean recognition problems,
including the example of computing N-bit parity with about N/2 sigmoids.

---------------------------------------------------------------------------
To obtain copies of the postscript file, please use Jordan Pollack's service:

Example:
unix> ftp cheops.cis.ohio-state.edu          # (or ftp 128.146.8.62)
Name (cheops.cis.ohio-state.edu:): anonymous
Password (cheops.cis.ohio-state.edu:anonymous): <ret>
ftp> cd pub/neuroprose
ftp> binary
ftp> get
(remote-file) sontag.capabilities.ps.Z
(local-file) foo.ps.Z
ftp> quit
unix> uncompress foo.ps
unix> lpr -P(your_local_postscript_printer) foo.ps

----------------------------------------------------------------------------
If you have any difficulties with the above, please send e-mail to
sontag at hilbert.rutgers.edu.   DO NOT "reply" to this message, please.