slehar at park.bu.edu (Steve Lehar) writes:
>Distributiveness is the way that neurons tend to branch out so
>abundantly, receiving input from and sending output to thousands of
>other cells. This is very different from computer systems, but we are
~~~~~~~~~~~
There is thousands of synapses, but several connect to same neurons?
Is there any litterature on this subject?
>The reason we abandoned it [analog computation] was because of
>it's complexity and chaotic tendancies.
No. The reason we abandoned analog computing was the evolution of
digital computers. With digital simulations of analogical signals we
could work faster and more accurately (at least in usual cases).
Actually neuron cell potentials are quantized in ion level. There is
abt 10^6 ions (abt 20 bits/cell) in one neural cell.
>The reason the brain uses it is because of it's complexity and
>chaotic tendancies.
Chaos and limit cycles compared to artificial networks (stable
attractors) is a very interesting subject.
>It is kind of like a system of balls connected by springs, where each
>spring represents a spatial constraint linking two balls. Short
>springs between nearby balls enforce local constraints, while long
>springs between whole groups of balls enforce more global constraints.
>Given certain inputs (some balls clamped into specific positions) the
>rest of the network will wiggle and jiggle until it finally relaxes
>into a global stability where the total energy of the system (sum of
>tensions on the springs) has reached a minimum.
This is how artificial neural nets work - not like the brain works.
Like you said earlier; the brain never converges.
What to do for artificial nets to get brain performance:
1. Good resources
(10^10 cells, 10^14 synapses)
2. Pulse code modulation
PCM makes it possible to self-organize simple and fast neural
structures to simple tasks and complex structures for complex
tasks.
3. Dynamic resource allocation (dynamic topology)
Neural structures are constructed during learning.
--> Neural environment will organize to do lazy computation.
This means the best result with the least computation.
Unnecessary neurons are killed.
4. Time-learning in neural level.
Each cell will delay the pulse; this would lead to better
time-learning with dynamic resource allocation.
5. Not to loose phase information
If any time-learning is implemented in artificial nets, it is
usually done syncronously; the net will lose phase information.
6. Time smearing of post synaptic potentials. (neurotransmitters)
If you use time smearing it is good to use PCM to save resources.
7. Dentritic computations.
Each neuron is like a hopfield-network (with only positive weights)
for its incoming synapses. (before the non-linear computation)
8. Pre-synaptic inhibiton.
This could be stupid to implement in artificial nets. It will
eat the resources and the results might not differ a lot.
But if you are planning to work on a biological style networks and
you have decided to use a big data structure for each synapse, why
not combine neurons and synapses. This would make it possible to
have pre-synaptic inhibition or any other synapse-synapse
connection.
I have been working with dynamic resource allocation, time-learning in
neural level and time smearing. I will try to implement PCM soon.
BTW, the estimations of the maximum physical limits for human brain
power are:
Energy consumption limit: 10^20 bits/s (this is absolute limit)
Cell energy limit: 10^9 bits/s (using the data of octobus neuron)
Cell energy limit: 10^11 bits/s (using the data of fly eye neuron)
Cell energy limit: 10^12 bits/s (using the data of bull frog neuron)
Synaptic bandwidth limit: 10^12 bits/s (estimating the brain as a 5 ms
discrete machine)
In these estimations I suppose there is no noise in the brain. I try
to find some time to write a proper paper on the methods...
--
more at stekt.oulu.fi - Jyrki Alakuijala - Atraintie 6, 90550 Oulu, Finland
+358-81542334, male, 20 years, University of Oulu, Neurosurgical Research Unit
