Modelling Symmetry Detection with Back-propagation Networks

[Cyril Latimer]

The following paper appeared in Spatial Vision, and reprints may be
requested from the address given below.

Latimer, C.R., Joung, W., & Stevens, C.J. Modelling symmetry detection with
back-propagation networks.  Spatial Vision, 1994, 8(4), 415-431.

Abstract

        This paper reports experimental data and results of network
simulations in a project on symmetry detection in small 6 x 6 binary
patterns.  Patterns were symmetrical about the vertical, horizontal,
positive-oblique or negative-oblique axis, and were viewed on a computer
screen.  Encouraged to react quickly and accurately, subjects indicated
axis of symmetry by pressing one of four designated keys.  Detection times
and errors were recorded.
        Back-propagation networks were trained to categorize the patterns
on the basis of axis of symmetry, and, by employing cascaded activation
functions on their output units, it was possible to compare network
performance with subjects' detection times.  Best  correspondence between
simulated and human detection-time functions was observed after the
networks had been given significantly more training  on patterns
symmetrical about the vertical and the horizontal axes.
        In comparison with no pre-training and pre-training with asymmetric
patterns, pre-training networks with sets of single vertical, horizontal,
positive-oblique or negative-oblique bars speeded subsequent learning  of
symmetrical patterns.  Results are discussed within the context of theories
suggesting that faster detection of symmetries about the vertical and
horizontal axes may be due to significantly more early experience with
stimuli oriented on these axes.

------------------------------- * ----------------------------------
Dr. Cyril R. Latimer                            Ph:  +61 2 351-2481
Department of Psychology     *     *            Fax: +61 2 351-2603
University of Sydney             *
NSW 2006, Australia                     email:  cyril at psych.su.oz.au

------------------------------ * -----------------------------------