Neuroprose paper: Representation of Temporal patterns in RNNs

[fred cummins]

           *** DO NOT FORWARD TO OTHER GROUPS ***


The following paper, to be presented as a poster at the 15th AGM of the
Cognitive Science Society, has been placed in the neuroprose archive.  FTP
instructions follow the abstract.

Fred Cummins
fcummins at cs.indiana.edu


__________________________________________________________________________

*** Representation of Temporal Patterns in Recurrent Neural Networks ***

To be presented at the 15th Annual Conference of the Cognitive Science Society

This paper investigates the way temporal patterns are represented by
recurrent networks. Small sequence discrimination tasks were devised
with the specific intention of requiring the network to base its
discrimination on the temporal structure of the input sequences.  The
resulting solutions  were then analyzed using elementary tools from dynamic
systems theory.  Our primary interest was to establish the relationship 
between the temporal characteristics of the training set and the representations
developed by the network.

Networks trained on a variety of simple sequence recognition tasks were
examined.  Analysis of the state space of unit activations allowed 
a direct view of the means employed by the network to solve
a given problem, and provided insight both into the class of
solutions these networks can produce and how these will
generalize to sequences outside the training set.  This intuitive
approach may be of help in assessing the potential of simply recurrent networks
for a variety of modelling problems.

__________________________________________________________________________

The paper is cummins.temporal-patterns.ps.Z in the neuroprose archives.

To retrieve this file from the neuroprose archives:

unix> ftp archive.cis.ohio-state.edu
Name (archive.cis.ohio-state.edu:whoever): anonymous
Password: (use your email address)
ftp> cd pub/neuroprose
ftp> binary
200 type set to I
ftp> get cummins.temporal-patterns.ps.Z
200 PORT command successful.
150 Opening BINARY mode data connection for rosen.advsim.ps.Z
.
ftp> quit
221 Goodbye.
unix> uncompress cummins.temporal-patterns.ps.Z
unix> lpr cummins.temporal-patterns.ps