levels

[Terence D. Sanger]

It seems to me that there are many different ways of describing any
phenomenon or algorithm in terms of levels.  "Levels of abstraction"
is probably only one, which might be interpreted to include both
biological hardware levels of organization (receptors, synapses, neurons,
Brodmann's areas) and the processing levels which a system goes through 
in order to interpret its environment (receptors, features, objects, 
interpretation).  As Sejnowski points out, Marr's levels 
(implementation, algorithm, theory) are an additional 
type.  Different concepts of level might have different theoretical uses, but
when it comes to trying to find examples in the hardware of a system,
(see Aaron Sloman's note) there may not be that many possibilities.

I would like to suggest another concept of level that might have
some basis in biological hardware.  I call it "anatomic levels".  The
idea is that the lower levels correspond to the local processing units
which (due to physical constraints) have access only to a small 
portion of the total number of inputs
and controls.  The higher levels progressively integrate input information
from lower levels and coordinate the outputs from lower levels.  An
example would be segments of a spinal cord performing maximal processing
based on the inputs and outputs available at that segment.  Propriospinal
communication would be the next level up, combining inputs and coordinating
motion between a few levels.  Sensory association cortex and perhaps 
supplementary motor cortex might (vaguely) correspond to higher levels
which integrate sensory information across modalities and coordinate
motor control across all spinal segments.

Note that relatively sophisticated processing (according to some measures)
might be occurring even within an individual spinal level.  "Complete"
interpretation of the available input and "optimal" control of the 
available outputs is theoretically possible, and might involve a good deal
of processing at multiple "levels of abstraction".

Does this have any relevance for computers?  Perhaps in a distributed 
processing system where nodes do not have access to complete sensory
information or control outputs, it would be useful to have an idea of
how to integrate sensory information across the entire network and how
to generate coordinated control.  

Terry Sanger
MIT E25-534
Cambridge, MA 02139
tds at ai.mit.edu