On Fri, Jun 26, 2015 at 12:10 AM, Mike Archbold <[email protected]> wrote:
>
> I've been trying to follow this thread on and off for days... it is
> hard to see what is being argued, let alone who is right!

I've been feeling the same way, and I started the thread!

When I started it, I simply hoped to point out (and get comments on) what I
believe to be a real possibility - namely that whole brain emulation (WBE),
one of the most promising and seemingly straightforward paths to AGI
(straightforward in the sense of requiring no major insights/breakthroughs
into the nature of intelligence, just lots of engineering), may not be so
straightforward after all. The reason I suggested was that even if we know
in principle how neurons interact, that interaction might not be
efficiently computable on any known computer hardware.

I made the analogy to how we know in principle how proteins fold into 3D
structures but can't in practice figure out even for moderate-length
protein sequences how they will actually fold-up into particular 3D
structures.

The reason protein folding is not efficiently computable (and is in fact,
NP-hard) is that the final 3D structure critically depends on interactions
between amino acids, and the number of these interactions that must be
accounted for grows exponentially with the number of amino acids in the
protein sequence.

I postulated that interactions in the brain that must be accounted for in
order to compute a future state of the brain from a given current state and
new (sensory) input could grow similarly rapidly (and therefore not be
efficiently computable) if the probability of a neuron firing depends not
just on discreet, graph-like patterns of connectivity via synapses (which
presumably could be efficiently emulated), but also on local potential
fields, which propagate independently of the connectivity pattern of
neurons (much like the electrostatic forces that govern interaction between
amino acids in protein folding).

These local potential fields are both created by, and have influence on,
the firing of (groups of) neurons, potentially creating a feedback cascade
effect that could be quite sensitive to small changes in the initial
configuration and the timing of active neurons. This sensitivity of the
final outcome (brain state) to small, non-local changes in the system's
initial conditions could make WBE intractable - even if we're one day able
to perfectly capture via new scanning methods the 3D structure of neurons
and their connectivity pattern.

The reason I think this interesting and relevant is that many high profile
people (e.g. singularity economist Robin Hanson) see WBE as the most likely
path to AGI because on the surface it seems like all that is required is
straightforward engineering - all we need is better scanning methods (e.g.
extending vitrification techniques already apparently quite good for mice
to work for human-sized brains), straightforward extension of current
neural modeling techniques (e.g. refinement to Hodgkin-Huxley model of
neurons) and more powerful computers on which to run the emulations.

If instead the WBE path to AGI proves very difficult or computationally
intractable, then alternative methods, like AGI based on cognitive
modeling, could be a better path to AGI.

I THINK the direction that this thread has taken since my original post was
towards a debate about whether certain forms of information (graphs vs.
topologies) can be efficiently modeled via data structures and computing
technology that exist.

My take on the issue is that if the answer you're trying to find depends on
interactions whose number grows exponentially with the number of elements
in the system you are modeling, than its pretty likely to be intractable
for systems of interesting size, like the brain, regardless of the clever
data structure you try to use. But I could be wrong, and my perspective may
assume that P != NP.

Furthermore, its an empirical question whether or not the number of
important interactions that must be modeled to accurately emulate the brain
is bound by the graph-like connectivity pattern of synapses between
neurons, or must include difficult-to-model field-like interactions between
neurons without direct synaptic connections between them, via  local
potential fields or something similar (e.g. non-targeted dispersion of
neuromodulators).

--Dean



-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com

Reply via email to