Biology

[Jim Bower]

	The current exchange concerning the neural basis of language processing 
reflects a general and perhaps growing tension in this field between
what is really known about biology, what is claimed to be known about biology 
(often in this field by those that don't actually do biology
but synthesize selected biological facts), and engineers 
interested in using the nervous system as a source of ideas for neural 
network implementations. A few comments seem appropriate:
 	First, there is absolutely no question that our real understanding of
how the nervous system works is extremely rudimentary.  This is as true
at the cognitive level as it is at the level of the neurobiological details.
If someone states otherwise they are probably selling something.  	
 	Second, understanding what is and is not known about biology requires
a considerable commitment to the study of biology itself.  Summary articles 
and general lectures at neural network conferences are not enough to
develop the intuition necessary to interpret neurobiological data.  This is
especially true in the case of lesion and psychophysical data which are in
any event problematically related to the actual structure of the brain.  
 	Third, while neurobiologists have and are continuing to collect massive
amounts of structural information about the nervous system, our ignorance 
is such that it is very difficult to even know where to begin in relating 
the abstract imaginings of neurologists, cognitive psychologists or
connectionists to neural structure.   Yet it is the firm belief of some of us
that the structure of the brain itself must guide these more abstract musings. 
Only hard work and cross training will allow this correspondence to
be made.  Non-biologists should also keep in mind that the lack of formalism 
in biology is not related exclusively to the inclinations of biologists.
It is also the case that we are studying the most complicated structures
known anywhere.  Physicists are still debating how to formally characterize 
the behavior of dripping faucets.
     With respect to the ongoing discussion of levels, for example, it is not
at all clear that feedforward networks of the connectionist type are even
a particularly appropriate metaphor for thinking about levels within the brain. 
This is especially true if a hierarchical organization is also implied.
Specifically, the usual description of a sensory to motor path within the
brain, with "lower levels"  of local sensory processing units feeding
"higher integrating levels" that in turn coordinate motor response is certainly 
a vast oversimplification and quite possibly conceptually wrong.  In
the case of visual processing, the often mentioned but still completely not
understood fact that there are 10 to 100 times more connections from the
visual cortex to the geniculate than vice versa at least obscures any simple
causal processing hierarchy.  Further, the sensory to motor, lower to higher 
to effector view of the brain would seem to completely break down when
one realizes that, under normal operating conditions (i.e., monkeys not in
chairs looking at television screens),  an animal itself controls the way it
seeks data.  This sensory acquisition process almost certainly reflects a
complex and evolving understanding of the object being explored.  Figuring
out how the deepest levels of the brain control sensory acquisition and
thus the neural flow of information through direct neural and indirect  
behavioral loops is likely to be an essential part of understanding how
brains operate in the world. 
 	Clearly, our understanding of how the nervous system works will not
only benefit from, but will be dependent on the fusion of computational
and neurobiological research.  However, any attempt to fake a fusion by
smoothing over the facts, and proceeding at full pace without concern for
the structural details of the nervous system itself, is likely to do more
harm than good.  

Jim Bower
Div. of Biology
Computational Neural Systems Program
Caltech