Artificial Neural Networks (ANNs) are useful and popular computational
but are they a suitable way to demonstrate the plausibility of neural
structures responsible for high-level human cognition? A new paper (PDF
format) by Peter R.
Krebs argues that ANNs are a bottom-up methodology and, as such, may
not be the best way up understanding the top-down concepts studied in
cognitive science. "When the tools (simple artificial neural
networks) to solve the problems (explaining aspects of higher cognitive
functions) are mismatched, models with little value in terms of
explaining functions of the human mind are produced." Though ANNs
are useful tools, they represent a universal framework for modeling any
cognitive theory. This means that modeling some aspect of human
cognition with an ANN merely shows that it's neurlogically possible but
cannot demonstrate that it is neurologically plausible.
I remember learning about ANNs at uni (my thesis actually involved
using ANNs in a basic classification task). Possibly tainted by my
previous, but limited, experience I have always looked at ANNs and the
such as useful tools to achieve an outcome, rather than as a method
for explaining the workings of the mind - to me, the ANN is a transfer
function you don't have to know - obviously better suited to LTI
systems than non-linear ones.
I remember that a friend's thesis involved trying to extract
classification algorithms out of a neural net (ie determine the
transfer function) - that was a harder job, and I can see how that
sort of work would be beneficial to trying to fathom how information
is stored in the brain - although Krebs has a point in his top-down
vice bottom-up view. After all, ANNs and SRNs are just models, and
there's no such thing as a universal model.
Nevertheless, I still would like to put an ANN into a mobile robot and
teach it that walls are bad, stairs are bad, low battery is bad,
people are good, recharge plate is good, etc. The problem is, that
ANNs generally need so, so much pre-processing (or a LOT of learning
time and CPU power).