Note to ZS readers: save this message for Monday, I just had to get it
off my HD tonight. (too lazy to hold it).
om
Today I've decided, foolishly, to spill the beans on some of my more
advanced concepts for AI. I'm going to cover the parts of my Hypermind
architecture that I can remember right now and discuss their feasibility
and the technical problems in implementing them on both hardware and as
a bio-retrofit.
The basic capabilities I'll be discussing are distinct but closely related:
-> Algorithmic learning.
-> multi-session intelligence with interleaved sleep and wake sessions.
-> multi-configuration intelligence
-> distributed intelligence
-> High performance intelligence.
om
One of the annoyingist things about learning is repetition, or having to
go through ginormous training sets to learn a simple idea. Neural nets,
both natural and artificial use what is called adaptive learning.
Obviously, it isn't all that great, even in the high performance neurons
we humans get to use. What is going on is that our populations of
neurons have to be slowly nudged into the shape we want them to be in.
This takes practice, repetition, and lots of Dr. Pepper. (not that I use
that specific drink).
What if we could learn algorithmically? What if we only needed to be
exposed to enough input to reasonably disambiguate the information to be
learned and then be able to use that knowledge, with masterful skill
until we run out of disk space and are forced to prune back the
less-useful stuff... The human memory system appears to be able to
buffer information in short term memory long enough for the long-term
neural nets to learn it, intellectually at least (with pathetically
limited bandwidth)... Skill memory is much trickier because it involves
wiring up what amount to new processor elements. So a student of math,
at any level, must work through many related problems before the skill
pathways begin to form, the long term memory is mostly useless for this
kind of skill... The amount of practice required to master a musical
instrument is legendary.
The first step is to identify what exactly needs to be learned. That is
why the human brain relies on its imagination as it's primary mode of
perception. It compares the actual scene with the imagined scene to
identify surprising stimuli which it has special neurons rigged to
detect and then focus as learning targets. The most annoying kinds of
sounds are sounds that can't be modeled by the brain and therefore are
constantly hitting this surprise circuit. The brain is certainly much
more advanced than current neural networks but can we go any further? We
need is a much more efficient means of generating motor programs. One of
the issues is that the brain is pattern based, not math based, a math
co-processor, much more precise and reliable than the cerebellum could
provide the means for much more precise mental rehearsals. Furthermore,
if the information is stored in discreet data structures rather than a
matrix-blob as current neural networks are (or as a physically fixed
cortical circuit), then an algebra and optimization process can be
performed on them to much more quickly derive a more optimal
representation that would otherwise take years to develop through practice.
Another problem with the matrix representation is that it simulates
neural plasticity with a vector of weights. In larger networks, this
vector is almost sure to be relatively sparse and therefore considerable
amounts of computation will be wasted computing N*0 . Therefore the
conventional systems are inherently unscaleable in addition to having
the previously mentioned weaknesses.
In a biological context we need pattern processing elements that can be
re-configured programatically and we need an organ that can inspect and
re-organize these pattern processing elements on a continual basis.
om
One of the crippling limitations of our current neural architecture is
that we can only do one thing at a time. The best we can do is rapidly
time-share our brains between two tasks. This will almost inevitably
lead to confusion as I frequently try to read an article and listen to a
video talk or internet radio talk show at the same time. Our brains also
must spend about 1/3rd of their time doing house-keeping tasks and
simulations to rehearse behaviors and consolidate memories. What if we
could treat our knowledge and memories as a mainframe hosted database
and run tens of thousands of fully isolated concurrent processes over
it. It might cut out useful forms of confusion but it would also allow
us to fire up as many sessions as we need to process our 21st century
information diet. This, again, requires us to be able to process
memories and skills as discreet units that can be treated symbolically
so that cached copies in different partitions can be re-integrated, as
it has been noted that this won't work in neural nets.
So we need a cognitive architecture that can be multi-tasked out to the
limit of the hardware... Conceptually, this is no different than how an
operating system carries out multitasking, it is only necessary to be
able to virtualize the hardware and run the virtual instances
concurrently. Then RDBMS techniques can be applied to the stored
abstractions and there you go...
In a biological context, it is necessary to be able to tag and index
activation patterns so that they can be processed independently against
the set of abstractions.
om
When one thinks about all the kinds of things one would want to extend
one's mind into in all of transhumandom (real world applications always
being more interesting than VR...) one is boggled by the variety. On the
extreme, you might want a highly optimized flame-bot for getting into
IRC and mailing list flame-wars, its modalities would be primarily
text-based and would only have enough visualization to be able to
properly interpret the language. On the other extreme you would might
have a very large and highly complex virtual body with many conventional
and unconventional modalities for 3rd generation cyber-sex... Both
systems would act concurrently on the same database... The basic problem
is that you need to select and configure an ad-hoc network of functional
units as needed for the selected embodyment. The fixed white-matter
pathways one finds in conventional brains (and simulations thereof)
hardly provides for any flexibility. The only solution is to chuck it
and use computer data structures.
A biologically embodied superintelligence wouldn't need to deal with as
much change as mentioned above. It would be sufficient that it merely
have a few spare general-purpose channels that can be tasked as
necessary. Additional capabilities would come through networking in
semi-sentient sub-units.
om
That brings us to distributed intelligence. For many reasons it is
desirable to spread a mind across geographically separated locations.
The physical security of any one site can never be perfect and it is
probable that low-latency interactions will be desirable in a number of
contexts. The first issue is latency, the second issue is the one Eugene
likes to beat us over the head with, synchronizing slower hosts with
faster hosts. There are two basic types of interactions here. The first
is the knowledge, skills, and memories. This level of knowledge, as I
proposed above, can be treated like a database and therefore the
standard synchronization techniques can be applied. Latency is not a
major problem though bandwidth is. As bandwidth becomes choked, each
side the link will have to prioritize traffic. This is not a problem at
all. What is a problem is the high level executive function, the thread
of consciousness who's modality is to manage all of the other parallel
operations. As the system becomes larger, it will be necessary to apply
a deeper and deeper hierarchy of executive functions. Unless there's a
major breakthrough in quantum communication, it appears that it will be
necessary that the top level executive function will necessarily be
quite slow. However, this does not place any constraints on the
subordinate exeucitives, or the workers. Unlike what you may have read
in sci-fi, the parts of the AI you actually interact with locally will
run at full speed but they will not have any special privileges in
communicating with other agents within the same mind on a distant planet.
Eugene says that neural interfaces are pointless because human neurology
is so many orders of magnitude slower than an AI substrate (actually, he
claims this is true of uploads too!!!). He never goes past the quick
put-down. He assumes what he perceives to be perfectly obvious to be an
insurmountable, unavoidable obstacle. The question is, exactly how much
communication do you need between the "fast" parts of your mind and the
legacy parts of your mind? Is there sufficient bandwidth available and
does the latency create a bottleneck?
As I explained above, when bandwidth is limited, then only the most
important messages are sent. This is an inherent limitation in the human
brain that simply cannot be overcome by any conceivable technique,
including uploading, because the neural architecture is too limited in
capacity and has the previously mentioned limitations in learning and
scalability. It is also acknowledged that a slow-time or even uploaded
human would have difficulty managing more than a small number of
interlinked agents. The only way around that is to use a more powerful
executive function external to the base neural architecture. This does
have the side effect of reducing you to an appendage of a larger being
but I'm pretty sure that's the best deal you can possibly get. The best
part of it, you get to live through the process!
om
Now for the last part of the discussion, pushing performance through the
roof. "clock speed" and even basic algorithmic tweaks are too trivial to
mention. Where the real gold is, is in expanding the architecture to be
able to grok a chemistry database of 2 * 10^6 unique chemicals. Or truly
understand 5-space physics. What you need is the ability to manipulate
and extend the architecture of the mind as the cognitive demands arise.
It's difficult to say when a new domain sub-processor would be
required... Maybe a frustration detector would be able to identify where
the existing system is failing, or maybe the problem would be optimized
out by assuming everything is higher-dimensional from the outset and
only reducing dimensions as the true eigen-spaces become apparent.
Regardless, the true power of superintelligence is the ability to expand
and reconfigure as the need arises. This approach, and only this
approach is capable of tackling the grand challenges of the universe.
Conclusion: superintelligence looks nothing like an upload! =P
--
E T F
N H E
D E D
Powers are not rights.
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968
Powered by Listbox: http://www.listbox.com
