[ExI] Brain emulation, regions and AGI [WAS Re: Kelly's future]

[Kelly Anderson]

On Mon, May 30, 2011 at 2:35 PM, Richard Loosemore <rpwl at lightlink.com> wrote:
> Kelly Anderson wrote:
>>
>> On Fri, May 27, 2011 at 7:18 AM, Richard Loosemore <rpwl at lightlink.com>
>> wrote:
>>>
>> Thanks. That helps me to understand some things better. My
>> understanding is that there are some divisions that are pretty
>> structural, such as the brain stem... but if I understand what you're
>> saying it is that in the neocortex, there aren't such sub organs.
>> Right?
>
> Hmmmm....  [ponders long and hard].  There are areas that are so domain
> specific, that some might call them sub-organs.  The visual areas at the
> back do a heck of a lot of processing that is the same in most individuals.

Yes, I've heard that before.

>  For example, there are separate dorsal and ventral pathways, which seem to
> split the visual processing into two paths, one of which computes
> vision-for-object-recognition, while the other computes vision-for-action.
>  (That is very approximate).  And there is an area of the left dorso-lateral
> prefrontal cortex (lDLPFC) that sometimes appears to be extremely specific
> (but to a cluster of different tasks.

OK.

> But I myself would not call these modules.  Partly that is because if you
> look at the wiring at the level of columns and microcolumns, you see very
> much the same patterns in those different places I just mentioned.  So, to
> my mind, what is happening is that the same basic style of architecture is
> being used (a large array of repeating units - the columns), but the types
> of "concepts" that accumulate in those repeating units, during development,
> end up being specialized.

Are there not different types of neurons in different areas of the
brain? Would this not contribute to different algorithms being
applied?

> The way I interpret the specialization is as follows.  There is incoming
> traffic to the cortex from many sources, some of which certainly do do some
> preprocessing.

Clearly, this is so.

> Now, those input wires arrive at the cortex at a bunch of
> places -- there is not one part of the cortext that acts as the gateway to
> the rest of the system, there are gateways scattered all over.  Also, there
> are some specific (probably hard-wired) superhighways that connect different
> parts of the cortext to one another.

Might these not be good "division" points when trying to do an
emulation of the whole?

> Now, with that combination of input
> ports and superhighways, the concepts that tend to be learned by the sea of
> cortical columns tend to be specialized for the same reason that the shops
> and businesses in a city tend to be specialized and localized .... because
> the wholesale vegetable market is located *here*, the cattle market *there*
> and the coffee houses *there* (so all the stock brokers arise in the coffee
> houses!), .. and so on.

It would be fascinating to understand how these neighborhoods get set
up. How much is genetic, how much experimental... probably a mix of
both, I would suspect.

> That picture I just gave means that all the columns have basically the same
> functionality, and any apparent organs are just the result of developmental
> pressures and some built in wiring.
>
>
> Well, from my perspective as a cognitive scientist/AGI person, I am not sure
> how valuable the low-level neuron stuff will be, in the end.

Just insofar as getting to the end, maybe.

> For example, I
> think we may have the functional model of a cortical column in the next five
> years, as a result of purely psychological analysis. That is, we could know
> pretty much what the columns are doing, without having to ask the
> neuroscientists what the circuitry looks like.  From the point of view of
> AGI, that would make us able to build a human-like AGI soon after, while the
> neuroscientists are still trying to figure out where to store the pictures
> of the trillions of brain slices they are collecting (never mind how to
> analyze those pictures).

Aren't there fully functional computational models of parts of the
brain now? Aren't those models based on bottom up analysis, rather
than top down?

> Did you see my conversation with Eugen last week, when I mentioned a picture
> of a neural area at one micron resolution (did you see the image I dropped
> on my website, with one pixel per micron?).  It is kind of fun to imagine
> the neuroscientists using an image like that to arrive at a sensible circuit
> diagram .... the pixels do not even show all of the smaller cell bodies,
> never mind all of the smaller wires.  And the synapses are one or two orders
> of magnitude below that level of detail.  So that is my reason for
> skepticism about WBE.  I am really puzzled that more people out there do not
> stop to think about it, before they jump on the WBE bandwagon.

The only reason I'm on the Whole Brain Emulation bandwagon is that the
brain is the only example of intelligence we have. I get the whole
bird/plane analogy, but without understanding the rules of avionics,
you can't build a plane, and studying birds is ONE WAY of
understanding avionics. Granted, this isn't the only way.

-Kelly