Selection process of synaptic connections

Matt Jones jonesmat at physiology.wisc.edu
Tue Nov 23 11:06:18 EST 2004


"Rich" <nomore at spam.no> wrote in message news:<Qzpnd.110312$bk1.81076 at fed1read05>...
> "Rich" <nomore at spam.no> wrote in message
> news:u7fnd.106357$bk1.57913 at fed1read05...
> > http://www.napa.ufl.edu/2004news/braindish.htm
> >

> >
> > 1. What is going on when a neuron "extends a process" (sampling
>  neighboring
> > neurons) before making its final connection? What exactly is it looking
>  for,
> > just the nearest neighbor? They describe it as though the connections are
> > searching for something in particular.

There are a variety of secreted molecules and cell-surface molecules
that collectively are known as "trophic factors" - i.e., molecular
signals that influence the outgrowth, retraction, direction of growth
and stability of developing axons/dendrites, etc.

A google or PubMed search for "trophic factors", "neurotrophins", or
"cell-surface adhesion molecules" will probably turn up enough
references to choke a horse.



> > 2. They have an array of electrodes connected to cells in a petri dish.
>  How
> > are the neurons "trained" to know right from wrong?

> > The neurons don't even know *what* they are analyzing, so how can they
> > "respond accordingly"? For the flight path of an airplane (for example),
>  how
> > do the cells know right from wrong or nose diving from flying straight...
> > What sort of training is going on??

There must be some sort of "reward/punishment" proceedure implemented
by the researchers. For example, whenever the "dishbrain" does
something bad (like crash the plane into a simultated crowd of
innocent bystanders or strafing a red cross tent) , perhaps they
reduce calcium in the dish to promote long-term depression of the
syanapses that were active when making that "decision". Bad dishbrain!
Bad! Bad!

In contrast, whenever it does something good (like bombing an enemy
bunker while incurring miniml collateral damage) perhaps they elevate
external calcium or apply tetanic stimulation to cause long-term
potentiation and thus strengthen the participating synapses. Good
dishbrain! Good boy!




> >
> > In artificial neurons (programmed in software or on hardware) synaptic
> > weights are adjusted by means of a training algorithm.. the contribution
>  of
> > error from each neuron is calculated and synaptic weight changes are made
> > afterwards in a backwards pass (back propagation). But what could be going
> > on with the experiment on real live cells? Just by sending an electrical
> > current as feedback, how do the neural cells in the petri dish know that
> > they have done something correct or incorrect in order to make their
> > adjustments?


My guess is that it's a supervised training algorithm, using a
standard Hebbian learning rule, something like I described above.
Also, I very much doubt whether you need real neurons to do this.
Chances are it would also work with an appropriate multilayer
artificial neural network.

Two more comments: 
1) It sure would be nice to see the data, instead of just the hype. 

2) Why a flight simulator, for heaven's sake? Why not sonmething truly
useful, like a robot that takes out the garbage, or makes coffee and
toast while you're in the shower?

Matt



More information about the Neur-sci mailing list