Xref: utzoo ont.events:1380 uw.talks:72 uw.cs.grad:61
Path: utzoo!utgpu!watmath!watdragon!ylkingsbury
From: ylkingsbury@watdragon.waterloo.edu (Yvonne Kingsbury)
Newsgroups: ont.events,uw.talks,uw.cs.grad
Subject: ICR Colloquium
Keywords: "Using Spatial Coherence as a Local Teacher for a Neural Network".
Message-ID: <18244@watdragon.waterloo.edu>
Date: 16 Nov 89 16:52:06 GMT
Distribution: ont
Organization: U of Waterloo, Ontario
Lines: 53



                 The University of Waterloo
                   200 University Avenue
                     Waterloo, Ontario


         The Institute of Computer Research (ICR)

                  Presents a Colloquium on


    Using Spatial Coherence as a Local Teacher for a Neural Network


by    Dr. Geoffrey Hinton

of    Dept. of Computer Science and Psychology, University of Toronto



DATE:      Wednesday, November 22, 1989
TIME:      3:30 p.m.
LOCATION:  William G. Davis Computer Research Centre, Room 1302


ABSTRACT

A  major  goal  of   research   on   unsupervised   learning
procedures   is  to   discover  an  objective  function that
defines the quality of an  internal  representation  without
any  externally  supplied   information  about  the  desired
outputs of the system.  If such a function could  be  found,
it  should  allow  a  hierarchy  of  representations  to  be
organized  bottom-up in a time roughly linear in  the  depth
of the network.  This would allow much faster learning  than
supervised procedures  which  are  generally  very  slow  in
networks with many layers of hidden units. We propose that a
good  objective  for  perceptual   learning  is  to  extract
higher-order  features  that  are  coherent  across  time or
space.  This can  be   done   by   maximizing  the  explicit
mutual   information   between   parameters  extracted  from
spatially or temporally adjacent   parts   of   the   input.
Recent  results  obtained  by Sue Becker show that this kind
of objective function can  be  used  to  discover  depth  in
random-dot  stereograms.   The   approach  can be applied to
many other types of unsupervised  perceptual  learning.   In
particular,  it  should  be able to discover the  underlying
three-dimensional  shapes  of  objects  when presented  with
an ensemble of two-dimensional images.


Everyone is welcome.    Refreshments served.