Question on BP..

[slehar@bucasb.BU.EDU]

Yeah- I did a hand-sketched character recognition program with
backprop (see ICNN 1986 was it? the Boston conference). I had a
100 x 100 sketch pad which I down-sampled to a 10 x 10 grid which
became an input layer of 100 neurons.  A run-time option allowed
for either thresholding, which produced on or off binary inputs,
or proportional sampling, which produced graded neurons between
0.0 and 1.0.  The algorithm quite effortlessly handles both.

To the human observer, the graded inputs looked much more 
recognizable because you could tell whether the sketch just
nicked the corner of a grid square, or ran smack through the
middle of it.  There was, in fact, more information in the graded
version.

To my surprise, backprop actually prefered the binarized version.
It learned faster, and discriminated better.