Starting program has goals.
Goals spin off sub goals by calling for any parallel
processes who can give answers.
Parallel processors are set in restrictive domains with
certain pre-programmed 'knowledge' and 'capabilities',
and
a set of external algorithms and programs that can be
run
on them to be loaded if needed, or self programmed if
no existing program is available.
Empty parallel processors cooperate in accepting a
needed
process by communicating between each other and if
they
are inside the correct "domain" they may program
themselves, using available knowledge from the
community or even the web, so that they can answer the
problem as best and fastest possible.
The system evolves and works in parallel. Different
versions of the system (at different evolutionary times,
or
evolved again at a different time or with different
conditions) could be compared to each other, choosing
hybrids of the system that best fit the problem and
desired
results.
The classical OCR paradigm would not be altogether
"neural network" algorithemless, but on the other hand
will
not be run by a central algorithm controlled program
which
sets up the whole thing and them accumulates results.
Rather, there will be competing but cooperating input
sources and competing but cooperating processing units,
and perhaps a few competing but cooperating output
venues, similar to our sentences which can be sometimes
constructed partially in advance and partially "on the run"
There will also be processes closing if things aren't
working correctly.