Connectionists: How the brain works

[Hans du Buf]

On 05/28/2014 01:27 PM, Juyang Weng wrote:
> Danko:
>
> What about a primate, an animal, and an insect. How many traverses 
> does each have?
>
John, Danko, all,
That's a good question but difficult to answer (at least for me).
How can certain raven solve 3-step puzzles for getting at some food?
They have never seen the puzzle, they look at the available pieces from 
different
points, and they have the solution. Are they consciously reasoning?
Or is it a low-level process in which the rewards of different combinations
are predicted and summed in order to predict the total reward in order 
to access
the food or not? Like chess players using a lot of knowledge, experience and
intuition, or a chess program that simply goes through all possible 
combinations
until a certain depth?
Like us, mens agitat molem, they are tool users. I would put them at our
level because, apart from using tools and solving puzzles, they are very 
social:
they can collaborate, they are inventive to hide food while observing 
the others,
they can fool others, and they can detect intentions from others. Three 
traverses?
So is an ant intelligent? Perhaps a little bit. But the colony is much more
intelligent.
Is a neuron intelligent? Without any doubt, because the many chemical 
reactions
are very complicated and intricate, and it has a function, like an ant. 
You don't
need to answer the question whether the neuron's colony is intelligent.

One of the questions was: is there one equation?
I have been reading a lot about simple reaction-diffusion equations. 
Murray's
two-tome book Mathematical Biology, explaining why some zebras are not
striped vertically but horizontally, or even group interactions. Turing 
instabilities,
spatial coherence resonance near pattern-forming instabilities, mostly 
systems
composed of only two components. I have been experimenting with 
Swift-Hohenberg,
where heat conduction in a fluid suddenly jumps to convection loops (if 
you are
interested, please see the very nice demos of Michael Cross at Caltech). 
I managed
to simulate the Brusselator, and belief me, even using Cross' notes it 
was not
easy to get a stable simulation because there is always one critical 
value where the
system jumps from one attractor to another. Only TWO components!

Nevertheless, we can use dynamic neural field (DNF) theory and 
attractors to track
multiple objects in complex scenes (vision) and to control limb joints 
for e.g. walking.
This is low-level. One level higher DNFs can be used to predict rewards 
and to
generate visuomotor sequences. A very complex system of massively 
parallel DNFs.
But it is still reactive.
The question, which I touched upon in a previous email, is whether the 
same principles
can be used to balance narrow and broad attention, both endogenous and 
exogenous.
To balance between excessive control and lack of control. I think this 
is possible.
But the ultimate question is whether it could be applied at the highest 
level.
This involves so many things. When I read a paper saying "this filter is 
implemented
by the following Gaussian" my brain expects to see two times sigma 
squared. This is
still experience and different memories. If you want to build an 
intelligent machine
which can solve a 3-step puzzle which it has never seen before, only 
with the goal
of satisfying a reward, you need a lot of experience and memories etc. 
before you
can even start to think about it. But, at the very end, the circuits 
which solve the
problem, and those which we think (!) are responsible for taking 
conscious decisions,
are all the same.
Do not expect a God in our machine, but: Deus ex machina.
Hmmm, reminds me of the fact that our brain still has a religious 
region, which
evolved in our ancestors to explain some natural phenomena and to 
establish a few
basic rules for the survival of the tribe.
Hans

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20140528/ffe83249/attachment.html>