HOW CAN RAPID CHANGES IN CONNECTIVITY IN NEURAL NETWORKS BE USED?

What follows is just me thinking aloud and may be full of junk.

On the other hand, the questions may be important, and if anyone has
answers I'd like to hear them.

Mike Denham (Plymouth University) recently drew my attention to a fact
which appears not to be widely known, namely that recent discoveries
indicate that not only levels of activation of neurons but also
*connection strengths* can change rapidly, e.g. in milliseconds.

I.e. such changes do not require long training processes which gradually
alter synaptic weights. So neural *architectures* can change very
rapidly.

This was confirmed in conversations at the recent Foresight workshop
with Pete Redgrave, a neuropsychologist at Sheffield University, and by
Walter Freeman (Berkeley University), whom I met at the ASSC7 conference
in Memphis a few weeks ago http://www.cs.memphis.edu/~assc7/

I have not yet read anything detailed on this, though I gave google
"brain connectivity change" and got this, among other things:

http://www.neuroinf.org/workshop/Cambridge-03-03/index.shtml
  Brain Connectivity Workshop
  May 1 - 3, 2003, Cambridge, England

The link on the web site to abstracts and presentations is broken, but
this works:
    http://www.fil.ion.ucl.ac.uk/resources/braincon/braincon.html

I did not find that anything there really answered my questions
about how such rapid changes in connectivity might be used, and what
mechanisms would make such uses possible.

NOTE:
If a neural net can change its architecture in milliseconds, by changing
its connectivity, I suspect this could be the basis of a very important
new kind of computational mechanism (or information-processing
mechanism, some might prefer to say).

For instance, until I learnt about the possibility of such rapid changes
in connectivity I was doubtful of the possibility of using neural nets
to explain things like our ability to do algebra rapidly in our heads,
or compose sentences like the one I am typing now, or understand
sentences with quite complex structures like the one you are reading
now, or the ability to rapidly take in a complex scene, e.g. on emerging
from the underground in the busy centre of an unfamiliar city.

I thought it obvious that reactive sub-mechanisms in a complete
architecture could make good uses of neural nets, e.g. mainly 'forward
chaining' neural nets (including loops). I also thought it obvious that
they could be useful for a large long-term associative memory used by a
deliberative mechanism (and the deliberative components of a
meta-management mechanism). But I did not think neural nets as currently
understood could be useful for mechanisms involving the kind of rapid
kind of structural reorganisation involved in high level perceptual
processes or in the rapidly changing constructions (in relatively short
term memory or memories - workspaces) of deliberative mechanisms.

That was because I thought structural changes in neural nets were
inherently very slow and required much repetitive training to alter
synaptic strengths. But if those strengths can vary considerably
within a few mechanisms, effectively producing rapid changes in neural
architectures, then that opens up new vistas.

QUESTIONS:
There are several questions about how rapidly changing connectivity in
neural nets might be useful:

1. What can such changeable connectivity be used to represent or
encode or control (i.e. what sorts of rapidly changing semantics
or control information could they express)?

    (There are many examples of changing trees and networks used in AI
    and computer science for many different purposes, e.g. representing
    partial parses, structural percepts, partial plans, ontologies,
    constraint propagation, state-transition systems, etc. Maybe there
    are lots more that we have not yet thought of.)

2. How can change in a network structure be used to initiate or control
or modulate processes?

Here are some kinds of answers, though no doubt there are many more:

 a. A structure can be "interpreted" by some process, e.g. something
    comparing two structures in order to build a description of their
    similarities and differences (as in perceiving structural
    relationships or analogies), or something interpreting a structure
    as a complex plan that can be executed, with conditional branches
    to cope with unknowns in the environment.

    It's not clear how such an interpreter could be implemented so as
    to take advantage of rapid changes in connectivity of a network.

    What could such an interpreter be in a brain? E.g. how would the
    changing structure be 'read' for use somewhere?
    Could this be part of the function of brain waves (something like
    reading memory contents in a computer??)
    Any ideas?

 b. The creation of the structure initiates or modifies other
    processes. E.g if impulses flowing through the network are used
    to control a large collection of other things, e.g. other neurons,
    or a collection of muscles, then changes of connectivity in the
    control network will alter the pattern of impulses delivered at
    any time. Perhaps this could control large collections of
    coordinated muscle movements, e.g jumping up to catch a swaying
    branch, or moving a hand carefully through a thorn-bush to pick
    a berry, or perhaps a bird manoeuvring a twig into a half-built
    nets.

 c. The 'interpretation' process is distributed over lots of local
    changes: if the connectivity in network N changes, that may be
    reflected in lots of local changes at various parts of N that
    influence other things, e.g. local chemical processes are changed,
    or other networks are changed by the fact that impulses flowing
    through various parts of N change when the connectivity changes.


3. How can the rapid changes in connectivity be produced in a
coordinated way, as might be required to achieve some high level goal
in a complex, structured, task domain?

Again, I can imagine various different mechanisms:

 a. Network N1 has synapses that are modified by signals coming
    through network N2, where some or all of the 'tips' of N2 are on the
    connections of N1, for instance, Changing patterns of neural
    impulses going through N2 will then alter the pattern of
    connectivity in N1. This would not be a good solution if neural
    propagation speeds are too slow to allow full advantage to be
    taken of the speed at which synapses can be made to change.
    There could also be synchronisation problems, if synchronous
    change is required. But it might be very useful for certain
    classes of tasks.
    (This is another view of answer (c) to question 2.)

 b. Chemical signals pumped through the relevant region of the
    brain might use molecules that turn various types of connections
    on and off. This might also be speed limited -- depending how fast
    the molecules can flow.

    There are also questions about the expressive power of a soup of
    molecules of various kinds. The mechanism might work well if there
    are K varieties of connection types and a mixture of up to K
    molecules could be used to determine which connections are on and
    which off. (Note the timing/synchronisation problems if chemicals
    move slowly through the brain.)

    NB: It would be harder to switch connection states on the basis of
    *relational* properties (e.g. turn on all switches required for
    understanding french, or for thinking about how to debug a recursive
    program, or ...). For typically the conditions would not be locally
    checkable unless some prior process had 'compiled' relational facts
    into local facts (e.g. first mark all individuals married to someone
    who can program computers).

 c. I don't know if synapses can respond to electromagnetic
    signals (if so what are we all doing to our brains nowadays, with
    mobile phones and other things?) but if they can then electrical
    brain waves might carry information about which connections should
    be turned on and which off. This could be seen as a potentially very
    powerful (?) generalisation of ways in which computer memories work?

    Of course this raises the problem of how the waves are generated
    and how the appropriate information is encoded in them.

    Also, if the wave patterns are bandwidth-limited then they may be
    able to control only a small subset of networks, or may be able only
    to do rather coarse-grained global control.

No doubt there are important questions not covered by the above three,
and classes of answers that I have not considered. I did not see any of
these issues addressed in my quick skim of the online abstracts of the
May conference.


One class of test cases for good theories would be the ability to
explain the speed at which we can cope with novel visual input, e.g.
coming out of a subway into the middle of a unfamiliar city, or watching
a troupe of ballet dancers, or sight-reading a bach fugue on the piano
or organ - which some people can do, though I can't.

The last example is particularly interesting because it involves very
rapid coordination of both structures involved in interpreting complex
visual input and structures involved in controlling complex actions with
a mixture of discrete and continuous features.

Many human and animal activities have 'slower' and 'less intricate'
versions of the same general nature, e.g. a child assembling a structure
from blocks: the unfinished tower and configuration or remaining blocks
plays the role of the musical notation on the sheet which has to be
'sight-read' as determining affordances for selecting new possible
actions, and also for controlling their execution.

If anyone reading this has any observations on the forms of information
processing that might be achieved by rapidly changing connectivity in
networks of networks of networks, please let me know!

IN CASE DYNAMICAL SYSTEMS THEORISTS READ THIS:
Obviously such a system can be construed as a dynamical system with a
*state* represented by a large collection of changing values (including
both connection strengths and activation levels, and maybe also lots
more detail, e.g. if signals do not travel instantaneously along neural
fibres) and with *behaviour laws* represented by large numbers of
differential equations.

But it's not obvious that such a 'dynamcal systems' view will be any
more informative than the (correct but very low level) dynamical systems
view of my computer running linux and the editor, mailer, web-browser,
spelling checker, I am using, among many other things.

Dynamical systems are needed as the infrastructure for all sorts of
virtual machines whose properties may be quite different from the
properties of the physical implementation.

I have the impression that a great deal of neuroscience (e.g. research
on brain imaging of various kinds) fails to take account of this, though
some researchers are well aware of the problem, e.g. in these slides by
Paul Fletcher, if I have not misunderstood them:
    http://www-bmu.psychiatry.cam.ac.uk/SEMINARS/store/20020307fletcher.ppt
    Maps and metaphors the perils of functional neuroimaging


I have raised the questions above as examples of apparently very
important issues that a 'grand challenge' project on 'Architecture of
Brain and Mind' would need to address.

They may have importance going way beyond such a project if they point
towards powerful forms of information-processing (computation) that
nobody has even begun to think about in computer science?

E.g. if some of the changes of connectivity are not discrete, but
inherently continuous, with non-linear feedback loops, then perhaps the
processes that arise *cannot* be modelled on Turing machines or the
various formal equivalents.

But Turing machines were not designed to enable a complex machine with
many moving parts to interact with a constantly changing, highly
structured environment.

(Many current AI robots don't do much of that either!)

Aaron