bradski@park.bu.edu (Gary Bradski) (03/22/91)
Decided to kick this idea around: When one sets out to create a machine that can operate as an independent agent in the world, one quickly finds that the problem is computationally impossible. [Where "quickly" can be on the order of 3 decades]. The usual response is to work on some toy problem and invoke the gods of massive parallelism to scale it up and do the rest. The gods of parallelism may help, but I think the demons of communication bottlenecks and "can't go faster than light" will still have the last laugh ['nother 3 decades...]. But -- all this time, even the smallest fly can navigate in three dimensions in a turbulent medium, find food, flee foes and mate with friends. I contend that the fly can do this not by doing *so* many computations, but because it is doing the *right* computations. I other words: the fastest, best generalizing, network/LISP code won't do very well if it's working on data that doesn't carry much, or obscures, the *information* of interest. I think the approach to take in developing intelligent machines is to first study of the form and content of the information that the environment provides. The best source on this is can be found in Ecological Psychology [James Gibson for founding/philosophy and works on perception, M. Turvey for more recent work]. Gibson's general idea is that the information needed to act in the world does not have to be synthesized -- it's already there. Gibson regarded perception as an active, searching process that extracts the embedded invariants (information) -- thus perception is considered as a full loop: the receptive organs as well as their motor and neural feedback tuning apparatus along with the intent to perceive. If the above view is correct, then the nervous system is relieved of an immense computational burden. AND, the types of machines built to operate in the world will concentrate not so much on computation as on extracting and tuning -- eg. coordinate transformations that bring out the embedded invariants and make them easily separable, correlation detectors, attentional and other types of focus etc... rather computations of masses, distances, reflectances, velocities, inertia's, inverse kinetics etc. I think the fly mentioned above speaks for this type of approach, the alternative is what I'll call the DARPA fly: 15 tons, 20 CRAYs and still can't do shit. >:-) -- Gary Bradski I'net: bradski@bucasb.bu.edu Center for Adaptive Systems Bitnet: bradski%thalamus@buacca Boston University. UUCP: {encore,harvard,uunet}!bu.edu!bradski 111 Cummington St, Boston MA 02215 I don't even agree with some of my opinions
pja@neuron.cis.ohio-state.edu (Peter J Angeline) (03/22/91)
In article <BRADSKI.91Mar21184156@park.bu.edu> bradski@park.bu.edu (Gary Bradski) writes:
If the [Ecological View of J. J. Gibson] is correct, then the nervous system
is relieved of an immense computational burden. AND, the types of machines
built to operate in the world will concentrate not so much on computation as
on extracting and tuning -- eg. coordinate transformations that bring out
the embedded invariants and make them easily separable, correlation
detectors, attentional and other types of focus etc... rather
computations of masses, distances, reflectances, velocities,
inertia's, inverse kinetics etc.
Wait, aren't the computations you're talking about really the same hard
computations that everyone is trying to do? Why is the nervous system relieved
of any computational burdens simply by recognizing that the information is
apparent in the "ambient array"? How are you planning to identify the
invarients in the environement if you're not going to concentrate on
"computation" just "extracting and tuning"? You've still got to build an "eye"
to get the invarients. Extraction of features is not that different than
extraction of invarients if you assume that the features are the invarients of
some environment. Choosing the right invarients IS NOT the hard problem in
perception, it's consistently acquiring the invarients in real time.
Gibson's view of perception is very important in that it specifys that the
"design" constraints of a biological perceptual system is hopelessly tied to
the invarients of the environment rather than being some idealized sensor. But
the hard computational problems are still there and are essentailly the same.
You've just narrowed the search a little, assuming you know the correct
invarients.
Gary Bradski I'net: bradski@bucasb.bu.edu
Center for Adaptive Systems Bitnet: bradski%thalamus@buacca
Boston University. UUCP: {encore,harvard,uunet}!bu.edu!bradski
111 Cummington St, Boston MA 02215
I don't even agree with some of my opinions
-pete angeline
--
-------------------------------------------------------------------------------
Peter J. Angeline ! Laboratory for AI Research (LAIR)
ARPA: ! THE Ohio State University, Columbus, Ohio 43210
pja@cis.ohio-state.edu ! "Nature is more ingenious than we are."
minsky@media-lab.MEDIA.MIT.EDU (Marvin Minsky) (03/24/91)
In article <PJA.91Mar22105320@neuron.cis.ohio-state.edu> pja@cis.ohio-state.edu writes: >In article <BRADSKI.91Mar21184156@park.bu.edu> bradski@park.bu.edu (Gary Bradski) writes: > If the [Ecological View of J. J. Gibson] is correct, then the nervous system > is relieved of an immense computational burden. AND, the types of machines > >Wait, aren't the computations you're talking about really the same hard >computations that everyone is trying to do? Why is the nervous system relieved >of any computational burdens simply by recognizing that the information is >apparent in the "ambient array"? How are you planning to identify the >invarients in the environement if you're not going to concentrate on >"computation" just "extracting and tuning"? > -pete angeline That sounds right -- and may be supported by considering how large are the brain regions involved with wisual processing; substantial portions of the posterior brain for sensory processing, and substantial portions of the anterior brain for oculomotor control -- and who knows how much more in between. Clearly the computational burden is very large. The question could be put the other way. Is the brain computing those invariants by brute force, because it has not evolved better ways? We won't know the answer, of course, until we have better theories of what computations are actually required for human vision -- and then, of bounding the requisite computation complexity.
uh311ae@sunmanager.lrz-muenchen.de (Henrik Klagges) (04/08/91)
Hello ! Thanks for kicking around ideas. Here they are back. Gary Bradski writes: |> The usual response is to work on some toy problem and invoke |> the gods of massive parallelism to scale it up and do the rest. Gods get their power from the number of their believers. From this I conclude that the GomP's are getting very strong ... |> I think the demons of communication bottlenecks Global communication, true. Local, wrong. No scaling problems, but of course limitations on algorithms. |> and "can't go faster than light" will still have the last laugh. 'Tragically, the early and, at its time, groundbraking species crayus ympsis fortranor couldn't adapt to this environmental constraint, died out and thus made space for connectius hillensis and its descendants.' [Van Nostrand's ectelopedian encyclopedia, 2211 AD, page ackermann(10) ] |> Gibson's general idea is that the information needed to act in the world |> does not have to be synthesized -- it's already there. The world is already there, too. |> Gibson regarded perception as an active, searching process that extracts |> the embedded invariants (information) -- thus perception is considered as |> a full loop: the receptive organs as well as their motor and neural feedback |> tuning apparatus along with the intent to perceive. (...) |> the nervous system is relieved of an immense computational burden. I like that you say it is all about processing the *right* information. I fear, however, that this often hard-wired (like in the early stages of visual processing) ability has been developed over millions of years of natural selection. I don't think drosophila's neurons do it right - I think drosophila's genome does it right. That means that marrying GA's with nets is the right direction. Cheers ! Rick@vee.lrz-muenchen.de Henrik Klagges, STM group at U of Munich (try also: uh311ae@sunmanager.lrz-muenchen.de, Compuserve: 100014,353) mung == mung until no good.