Re: Paths to Uploading

Anders Sandberg (asa@nada.kth.se)
05 Jan 1999 23:56:25 +0100

"Billy Brown" <bbrown@conemsco.com> writes:

> Well, I didn't say we could do it *now*. A MD-level sim of a few hundred
> cells would take, what, maybe 10^20 MIPS? I was just pointing out that even
> that is a far cry from a full-brain sim.

A molecular simulation is probably utterly beyond us right now, we can hardly simulate a single protein, let alone a cell. But I doubt that is necessary, if single molecules mattered so much we would be wildly unstable for the least disturbance. In an ordinary action potential there are around 10^5 ions jumping around, that is likely the minimal scale necessary (modulo certain highly significant molecules like PAF).

A fair cell simulation today eats perhaps 10 MIPS, but it depends on the level of detail. The scale is a few first order (somewhat stiff) differential equations per isopotential compartment; a real neuron would likely have many thousands of compartments but it is unclear how important the exact solution is.

> > Can we simplify neurons? Probably; this is just how nature made them,
> > and I think it is possible to create more computationally efficient
> > input-output mappings retaining the same properties. This is an
> > interesting problem, both from a research, a theoretical and an
> > uploading standpoint. What can we do away with without losing the
> > important? How much detail is needed? Could you upload yourself into
> > an integrate-and-fire model? Could this process be continued on higher
> > scales?
>
> I would think so. At the very least, you should be able to model the
> computation each neuron performs and do away with all the chemical details
> of how it gets done.

Exactly, although this is non-trivial and suggests the tricky questions about whether conscious experience is changed if you transform the computations into something equivalent (I would say no, but then again I'm a functionalist).

> If you want to get ambitious, you could probably also
> replace a large percentage of the simulated brain with ordinary software.
> Computer vision, hearing, and other sensory abilities will probably be much
> better than those of humans by then, for example. Of course, setting up the
> interface would be a bit of a challenge..

Yes, and our higher level systems reach in into the low level systems at a very early stage and interact with the "raw" data. In the lateral geniculate nucleus, for example, the visual signal from the retina gets mixed with signals from the higher visual cortices and *then* moves to the primary visual cortex.

> > What we need now is better data, better models and better computers.
>
> Amen.

Amen.

(you can put tax-deductible contributions into the bowl near the exits :-)

-- 
-----------------------------------------------------------------------
Anders Sandberg                                      Towards Ascension!
asa@nada.kth.se                            http://www.nada.kth.se/~asa/
GCS/M/S/O d++ -p+ c++++ !l u+ e++ m++ s+/+ n--- h+/* f+ g+ w++ t+ r+ !y