[MURG] - emulation difficulties

[Joseph J. Strout]

At 9:35 AM -0700 9/12/03, digfarenough wrote:

>I think, then, that it's safe to say we'll need to functionally 
>reproduce the neurons themselves in a network, to emulate a mind.

Yes.

>Two ways of doing this come to mind. One is to make neuronal models 
>of each cell type and use the Hodgkin-Huxley equations to calculate 
>the neuronal dynamics.

Well, neuroscience has advanced quite a bit since Hodgkin and Huxley; 
neuron models today are quite a bit more complex than that.  But this 
is the right basic idea.

>  The problem here, is that the HH equations are differential 
>equations, so that no matter how tiny a timestep you choose, 
>eventually the errors will greatly grow and the network will fall 
>apart.

No it won't.  If your integration method is poor and your time step 
is too large, then yes, the system will become unstable and fly apart 
(this is always fun to watch if you're doing a physics simulation!). 
But this is a problem that has long been solved.  With even a poor 
integration method, you can always take time steps sufficiently small 
that it remains stable; and there are more advanced integration 
methods that allow you to take a much larger timestep.

>  I don't know of any way of getting around this growing error in the system.

Error is different from instability.  It's unlikely that minor errors 
in the simulation will have any noticeable impact on the working of 
the mind.  Neurons work in a very chaotic environment, and if they 
were sensitive to small errors, the whole system would fall apart. 
Instead, there are numerous feedback loops everywhere you look, which 
result in the emergent behavior being very robust to a wide variety 
of influences.  Only extreme, systematic errors caused by things such 
as drugs or ischemia are able to disrupt the system behavior.

>This makes me believe that this method of emulation will only 
>provide a short life-span for an uploadee.

No, the uploadee will live just fine with this method.

>Another way of doing this (which may reduce to the above case, maybe 
>a TOE will help out here) is to simulate at a very low level 
>(atomic, perhaps subatomic) all the goings-ons in each neuron. If 
>you individually know the state of each voltage-dependant channel 
>and the location of every ion, you don't need to use diff eqs to 
>calculate them and can bypass the error.

That's not true; even the equations governing individual ion channels 
are differential.  So are individual atoms, for that matter. 
Basically, all the universe works on differential equations.  But 
this is not a serious problem.

>Does anyone know of other ways emulation has been proposed? Ways 
>that may actually work?

The first way you outlined above (with the caveat that an accurate 
compartmental model is a great deal more complex than just the HH 
equations) will work.

If you want to experiment with this sort of thing yourself, check out 
one of the neuron modeling libraries such as GENESIS, NEURON, or my 
own CONICAL.

Cheers,
- Joe

-- 
,------------------------------------------------------------------.
|    Joseph J. Strout         Check out the Mac Web Directory:     |
|    joe at strout.net           http://www.macwebdir.com/            |
`------------------------------------------------------------------'