X-Message-Number: 4398
Date: Mon, 15 May 1995 18:49:16 +0200 (MET DST)
From: Eugen Leitl <>
Subject: Uploading, yet again.
> Message #4392
> Date: Fri, 12 May 1995 07:32:52 -0700 (PDT)
> From: "Joseph J. Strout" <>
> Subject: Data required for uploading
[my original remark on neuron representation deleted]
>
> I suspect that this is highly oversimplistic. To configurate an accurate
> neural emulator, you need at least (a) detailed morphological data, and
> (b) information on the distribution of ion channels on the cell
> membranes. At least, this is the information needed by current cell-
> level simulators. It may be possible to reduce this requirement somewhat
I do not think we have to model the individual switching of ione
channels on nanometer scale. We are completely lost if it comes
to that. But we need an exceedingly accurate
mapping of the neuron bulk/synapse/postsynaptic membrane, may be
as high as at individual protein molecule level. This raw bulk
data, voxelspace coded, will be subject to different stages of
filtering in the signal processing pipeline. Until we arrive
at the weighted graph representation of the original minuscule
block of vitrified tissue together with representations of the
flat facets of it. Once one by one the whole of the frozen
brain lump is gone and the data has been uploaded into our
hypothetical neuro engine:
I think an _accurate_ emulation of the mind can be achieved by
representing the firing frequency by a small binary integer (about 6-8
bits per neuron body), may be a signal propagation delay (actually, this
might not be necessary, since hypergrid router does introduce delays
inherently) and a synapse weight (again, 6-8 bits per synapse). I think
at our current/near future integration density we can model about
1 cmm of the neocortex at realtime or an order of magnitude faster.
I will make an even bolder statement. I think there is a sufficiently
accurate mapping of bioNNs to any edge-of-chaos engine (I will repost
my Random Ramblings II which ricochetted off list to oversize which
might make matters clearer). I think this engine can be implemented
with current semiconductor photolithography technology. May be
with molecular circuits in the future. And this engine will be
computationally equivalent (homology at higher scale, as state
space/attractor kinetics) to our initial selves. It will be
probably be an automaton network (AN) engine.
> *if we can also take physiological measurements*, but this remains to be
> shown, and besides, you can get physiological measurements from dead
> tissue. But I wholly agree with your success criterion:
>
> > Provided I/my friends do not notice the difference between my original
> > deceased me and the simulation it is accurate. At least accurate enough
> > for me.
>
> You then tease us with:
[remark on nondestructive SQUID scan omitted]
>
> Please please please take the time to dig up a reference or two for us.
> This could be very important, but we can't evaluate it properly without
> seeing some original papers.
First:
I still believe into abrasional STM scan of immunolabeled vitrified
tissue. SQUIDs are a vague possibility, at best.
Details:
These SQUDS are simple superconducting devices with an additional
coil to adjust local magnetic field. Due to the tiny geometry there is
a high local magnetic field. Any small increase of it will cause the
superconductor to quench, which raises resistance by many orders of
magnitude. You can scan a dollar bill due to it's ferromagnetic
pigment at some 100 dpi resolution with a SQUID. Great picture.
I wish neurones were ferromagnetic, <sigh>.
BioNNs produce spatiotemporal patterns of electromagnetic
fields. A EM field of many firing neurons is simultaneously recorded
by an integrated SQUID array (there aren't integrated, yet. Novel
mercury-based high-temp superconductors are used to develop them
can't recall the ref, but it is a 1995 paper). This
interferogram of many neurons at several detector plates is the snapshot
of current activity in the brain. Using basically Fourier
Transform maths the original configuration can be reconstructed.
We retrieve the original patterns from its hologram interference
pattern. Using many snapshots the entire brain activity movie can be
decyphered.
Now for the bad news: the skull shields quite. Outer EM noise
exists. Both temporal and spatial resolution of the detectors
are limited. There is noise. There are sampling artefacts.
This means you'll probably never be able to single out
the contribution of a individual neuron from the chorus.
But it may not be needed. Since SQUID interferometry has the
best theoretically possible resolution of an in vivo scan
(far beyond PET or NMR imaging) it may suffice to record
and resolve spatiotemporal action of these systems, call
them agents, whatever. Using stupendous amounts of information
processing, one might be able generate a fuzzy
model of a person. How fuzzy I don't know. Probably too fuzzy
to be of any value. Future will show.
>
> -- Joe
>
> ,------------------------------------------------------------------.
> | Joseph J. Strout Department of Neuroscience, UCSD |
> | http://sdcc3.ucsd.edu/~jstrout/ |
> `------------------------------------------------------------------'
>
> Message #4394
> From: (Thomas Donaldson)
> Subject: NNs and our brain (and brains of salamanders)
> Date: Fri, 12 May 1995 12:59:32 -0700 (PDT)
>
> Mr. Eugen Leitl recently made a post commenting on what I had to say both
> about brains and about the powers or lack of powers of animals such as
> salamanders to recover memories after brain damage.
>
> While I've been interested in NN as computing devices for some time, too, I
was
> speaking in terms of the neurobiology of the subject. In this case, the
> "strength" of connections which is so important to the operation of a neural
> net SEEMS to be implemented by having multiple connections between the same
> two neurons. That is why connectivity is very important. If anyone wants a
If this is true, then connectivity is crucial. But current model
seem to imply that's all in the synapse. Now if the synapse just a
ternary element: a) it's not there b) its inhibitory c) it's excitatry
our scanning task becomes easy. Unfortunately I don't think this
will suffice.
> reasonable summary of these ideas as applied to real brains, they might read
> Steven Rose's book THE MAKING OF MEMORY. One of the very simplest observations
> and one known for a long time to support this is the fact that rats that
> learn a lot have a far more extensive neuropil than those that do not.
This is not a proof. There is a generic synapse proliferation, but this
is not a sufficient prerequisite for synapse/absense coding. It can
be just an increase in connectivity to anticipate future need. Or whatever.
But synapse strength modifications on their own are on record. They are real.
[snip]
> As for storing people on tapes, yes, that would work fine. Ideally, of course,
> you want some storage medium that deteriorates very slowly on a scale of
> centuries. If and when we can read people off into a computer, a lot of work
> will be done on finding a medium which does that --- right now I doubt it is
> true for anything but writing on paper or engraving on stone.
Not needed. You just use redundant encodings and copy the tapes
every 20 years or so. The information will remain intact for any length
of time provided you have infinite supply of fresh tapes. Join 3M.
> Finally, if Mr. Leitl believes that reading out the structure and connectivity
> of a brain despite freezing damage is now possible, I would like to know much
> more about his specific implementation of this idea. If he does not have one,
> then he should label his comment as a theoretical speculation.
Of course this is a theoretical speculation. If I had an implementation
I would have a cryonics corp of my own. But the ideas are straightforward:
+-------------------+
| block of vitrified|
| tissue |
+-------------------+
| destructive scan of sufficient resolution
|
+-------------------+
| voxel block |
+-------------------+
| noise reduction, contrast enhancement
| edge detection, further postprocessing
+-------------------+
| processed voxblock|
+-------------------+
| freeze artefact filter (an automaton
| network) trained to regenerate
| the original tissue. Based on optical
| (immunofluoro) microscopy and force
|
+-------------------+
| phoenixed voxblock|
+-------------------+
| membrane tracer pass
| synapse reader pass
| facet encoding
| compression
+-------------------+
| backup self data |
+-------------------+
| transfer/upload
+-------------------+
| simulation of the |
| personality model |
+-------------------+
Once we have a voxblock of sufficient resolution and
sufficient information (hence antibody labeling) we
have essentially solved the problem. Building the
simulation engine is the other big one left.
Don't get me wrong, these above was a simplified diagram.
Uploading is a monstrous project, far beyond the Manhattan
or Apollo or Human Genome project. While certain aspects of
it can be derived from other diciplines the final stages
will need heavy public funding. I don't see how we can
accomplish that.
Eugene.
Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=4398