To: Daniel Crespin, dcrespin at gauss.ciens.ucv.ve
These comments are made to the post you submitted by the above title on
23 May 1996 19:05:35 +0100. I have refrained from including that post
here. You said:
># Comment (3 of 5) to C: ....snip.... There does not seem to be an
>obvious criterion to tell when the extra examples are waste.
Three obvious criteria to tell "when the extra examples are wasted" (for
the example pattern classification problem posed herein):
a) Replicates of examples.
b) Examples which over-define boundary conditions. It takes only two
points to define a linear boundary condition. Two boundaries may share a
common boundary-specification-example. Any examples which fall on a
boundary once it is defined are superfluous.
c) Interior points. Once all boundaries of the polygon is defined,
interior points are superfluous.
Now, there seems to be angst expressed here towards algorithm designers
for their perceived abandonment of the constraint to honor
true-biological-rules during algorithm design. However, real brains work
mostly in the frequency domain space. Yet, the arguments here seem to
assume non-frequency-based NN i/o algorithms as the only extant
simulation modes. Shall the designers of mere pragmatic, engineering-like
algorithms be here judged by a standard to which this research has yet to
aspire?
Does the angst here stem from a bio-purist attitude? If so, it seems
discriminatory and expresses a degree of bio-elitism which is frankly
surprising.
If the goal of this work is modeling of biophysical processes utilized
during human experiences, then the purist approach may be well advised.
"Toy" applications such as presented here in the simple A vs B class
discrimination task are well posed mathematically. Yet to generalize such
simplistic mathematical treatment to all purviews of simulated neural
network applications and thereby also require all generalizations to
adhere strictly to supposed biophysical rules misses the pragmatic mark.
The simplest implementation of a simulated neural network system may only
need to be _inspired_ by an observation of biophysical reality; it need
not be a true emulation of bio-reality in order to be useful. Perhaps a
name change will suffice to separate the wheat from the chaff? Perhaps
"Simulated Pseudo-Neural Networks" or somesuch?
I fully expect that whatever is learned at _every_ stage of the research
to which Dr. Roy eludes will in some way find application to some
stand-alone or semi-conventional processing system of some type. The
ideas and discoveries will be used in every pragmatic way possible to
produce whatever is desired, and whatever is important enough to warrant
the effort. As for myself, I await the application of Dr. Roy's ideas to
the frequency-domain reality of neuron transfer functions and the
implicit, manifold methods of information processing effected by the
excitation and inhibition mechanisms composed in the "ionic soup" which
fills all the void space between neurons, synapses, dendrites, et al. In
the mean time, it's back to the application of simulated pseudo-neural
memories, pattern classifiers, optimizers, and hybrid systems to my
visually-driven and memory intensive problem domain.
Cheers,
-Jeff Baldwin
Mind&Vision Computer Systems