Full-text

URL: http://www.csee.uq.edu.au/~btonkes/point.ps.gz

Cached: PDF-245K    PS-320K    PS.gz-100K

SAVE AS an easy-to-recall long filename:

Filename format: author--year--title   PDF-245K    PS-100K    :: About GZip'd PS

Filename format: author--year--title--journal|proceedings|...--pages   PDF-245K    PS-100K

Related links

CiteSeer: http://citeseer.ist.psu.edu/tonkes99getting.html

Web search: Google Web Search   ::   Google Scholar

Within this site: References (21)

Paper at a Glance

Getting the Point Across: The Eect of Recurrent
Network Biases on the Evolution of a Simple
Language
Bradley Tonkes  Alan Blair 
Janet Wiles y
 Department of Computer Science and Electrical Engineering
y School of Psychology
University of Queensland, 4072
Queensland, Australia
May 19, 1999
Running Heading: \Getting the Point Across".
Keywords: Recurrent Neural Networks, Evolution of Communication
Correspondence should be addressed to:
Bradley Tonkes
Dept of Computer Science and Electrical Engineering
University of Queensland, 4072
Queensland, Australia
Phone: +61 7 3365 1157
email: btonkes@csee.uq.edu.au
1

Abstract Although language is an ostensibly symbolic medium, it is an open question whether the underlying concepts communicated through lan- guage are also symbolic, or whether they are ultimately more contin- uous and sub-symbolic in nature. Does language translate, in a sense, between a continuous concept space and a discrete, symbolic message space? If so, how could such a system have evolved? The languages which ultimately emerge will likely be inuenced by the constraints and biases of the communicating agents. We present a framework for exploring these issues, in which two re- current neural networks { playing the roles of sender and receiver { are able to develop their own language for communicating a set of concepts. Each concept, represented by a point in Euclidean space, is translated into a symbol sequence by an encoder network. This sequence is then serially transmitted to a decoder network which attempts to translate it back to the original concept. A series of simulations demonstrates that, within this framework, the biases of the sender are in conict with those of the receiver, and the resulting languages represent a compromise between these competing interests. 2 
1 Introduction Human languages provide a system for communicating sophisticated con- cepts that is unparalleled in the biological world. It is a puzzling question how these ...

@techreport{tonkes98gettingThe,
  author={Bradley Tonkes},
  title={Getting the Point Across: The Effect of Recurrent Network Biases on the Evolution of a Simple Language},
  year={1998},
  institution={},
  note={Extended version of SEAL98 paper, 26 pages.},
  url={http://www.isrl.uiuc.edu/~amag/langev/paper/tonkes98gettingThe.html}
}

HOME   ::  Technical Report List   ::   Technical Report	Comments to: junwang4 you-know-at gmail.com	Last update: 10/27/09