[Comp-neuro] Re: Attractors, variability and noise

Andrew Coward andrew.coward at anu.edu.au
Thu Aug 14 02:36:35 CEST 2008


On 13-Aug-08, at 3:36 PM, Prof Leslie Smith wrote:



Andrew Coward wrote

Understanding the brain will require an analogous hierarchy of  
descriptions. In other words, we have to find good approximations at  
several intermediate levels that can be mapped both into psychology  
and into more detailed neuron type models. One attempt at such a  
multilevel theory is [Coward 2005]. The danger of massive neuron  
modelling efforts is that we can create a system which may have  
properties similar to brains but do not help with genuine  
intellectual understanding.



Prof Leslie Smith wrote:

Engineers build complex system using hierarchical design: without  
question this makes them much easier to design and debug and extend  
etc. But biological systems are not designed as such: as such they  
don't need to work within hierarchies. Entities within biological  
systems can interact across levels. And they do: even simple GA based  
experiments like Adrian Thompson's FPGA's took advantage of  
unexpected interactions on an FPGA, and (for example) animal brains  
often use neuromodulators whose release affects a volume of the  
brain. Clearly, crossing levels is dangerous (which is why engineers  
avoid it) but it can result in highly efficient "clever" solutions  
(which is why evolution can utilise them).

A multi-level theory may help us to take an engineer's view of neural  
systems, but the view that it results in is likely to be incomplete.



No question that design “across levels” can have efficiency  
advantages. However, such advantages have to be “paid for” whether in  
an electronic system or a biological system.

In an electronic system, across levels design makes any future  
feature changes much more difficult. I had some experience with an  
operating system which took advantage of some timing relationships  
within one of its custom integrated circuits. Updating the integrated  
circuits as technology evolved became a horrendously complex  
undertaking. If across level design occurs in software, the result is  
that any future changes have large number of undesirable side  
effects, and the correction of these side effects introduces yet more  
cross level design fixes, making the next change even more difficult.  
Sometimes the only solution is to duplicate large amounts of code in  
the software system: one copy is changed for the new features, the  
other copy left unchanged for older features. In other words, cross  
level design when combined with a need to change has to be paid for  
with a large increase in system resources. This obviously results in  
increases in the size of the software, and I suspect that the huge  
increases on PC operating systems over the years may be partially a  
result of this type of problem.

However, biological systems are not immune to this type of problem.  
They need to change in the course of evolution, and some biological  
systems need to learn. The price of cross level “design” will be much  
the same, a much higher resource cost imposed on change and learning.  
Hence given resource limitations there will be strong natural  
selection advantages in favour of modular hierarchies. Obviously this  
is not an absolute, sometimes the competitive advantage of a “cross  
level” option will outweigh the resource penalties, but overall the  
modular hierarchies will be tend to be relatively “clean” because a  
brain that can perform a given set of functions with fewer resources  
will have a big selection advantage.

  If this were not the case, our brains could be “in principle”  
incomprehensible to our own intellectual capabilities, a position  
which has found favour with some philosophers. Hence the idea that  
natural selection tends to constrain brains into relatively clean  
modular hierachies is a relatively optimistic one.



Andrew Coward
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.neuroinf.org/pipermail/comp-neuro/attachments/20080813/1e04e13b/attachment.html


More information about the Comp-neuro mailing list