As somebody who (used to be) in the field of self-organizing systems, that is mostly true - we just have no good idea what the general principles are that are needed for an intelligent system.
There does seem to be some effect of quantity with respect to intelligence, however, for instance if you compare a bonobo and a human being.
In particular, I believe that, to a large part, intelligence requires the use of experimentation to develop. Thus, you need sensors _and_ actuators
CF: With respect to AI, there seem to be two main camps. One wants to emulate the behavior of intelligent beings, the other wants to emulate the biophysics. It's not clear that either approach will result in any huge breakthroughs in understanding consciousness.
GS: This does not strike me as accurate, or useful. Any AI must simulate animal behavior, at least in the same sense the flight of planes achieves some of the properties of flight in animals. A plane does not, as we have heard ad naseum, flap its wings, but it does go long distances without touching the ground, and it is not merely gliding. [Of course, interestingly, when we hear how AI might not resemble "natural intelligence" and are fed the plane/bird argument, we are not told that it is the case that the flight of the glider and the "flight" of the first animals DO closely resemble each other]. Is a connectionistic approach the former or the latter? I'm guessing that you would say that it is the latter (i.e., emulate biophysics), but this shows the conceptual error I am pointing out. If the NN does not emulate some functions of animals, then in what sense is it "AI." Incidentally - or maybe not so incidentally - a big part of the problem is the term "intelligence." One should consider that the term does not identify anything useful about behavior OR the brain.
CF: There was some promise in the careful study of the effects of certain substances on perception and consciousness, but this research was largely suppressed for political reasons around
1970 and nearly completely so later on for reasons which were and remain largely specious.
GS: I'm not sure what you are talking about here, but it can't be the field that continues to investigate the effects of drugs on behavior (i.e., behavioral pharmacology).
I don't think you can easily distinguish primates's brains from different species just looking under the microscope...I believe that you'd have difficulty even with rat and human tissue. That's the whole point of mammalian neocortex: it's the quitessential information processing engine, as far as we can tell, and it's only the way it's wired to peripherals, as well as its interactions with its peers (other parts of neocortex) that seem to determine its function.
Just so.
Well, you could do this in a simulated world, I suppose. Then you might run into the problem one of the guys from Thinking Machines had. It apears he set up a simulation for an artificial life model as a test for the CM-5 (IIRC). He said he debugged his physical world simulation by looking at the creatures that evolved. For instance, some of his first versions didn't quite get the principles of thermodynamics correct, so that some of the "animals" learnt to turn in a circle to gain energy for a jump 8-).
I am not sure how to apply the concept "distant interconnections" to the concept of "thousands of neurons." Are you calling a chain of local interconnections a distant interconnection? That's not the usage that computer scientists are taling about when they talk about distant interconnections.
If you apply the definition of "distant interconnections" used in computers to the brain, ther aren't any. Most neurons have an axon tree that is less than 1 or 2 nM across[1]. Although some partly myelinated axons can extend several millimeters in the neocortex and hippocampus, most neurons only have somewhat local connections.
References:
[1] Buhl EH, Halasy K, Somogyi P Diverse sources of hippocampal unitary inhibitory postsynaptic potentials and the number of synaptic release sites. Nature 368:823-828 (1994)
Sik A, Penttonen M, Ylinen A, Buzski G Hippocampal CA1 interneurons: an in vivo intracellular labeling study. Journal of Neuroscience 15:6651-6665 (1995)
[2] Sik A, Ylinen A, Penttonen M, Buzski G, Inhibitory CA1-CA3-hilar region feedback in the hippocampus. Science 265:1722-1724 (1994)
[3] Kisvarday ZF, Beaulieu C, Eysel UT, Network of GABAergic large basket cells in cat visual cortex (area 18): implication for lateral disinhibition. Journal of Computational Neurology 327:398-415 (1993)
Gupta A, Wang Y, Markram H, Organizing principles for a diversity of GABAergic interneurons and synapses in the neocortex. Science 287:273-278 (2000)
Salin PA, Prince DA, Electrophysiological mapping of GABAA receptor- mediated inhibition in adult rat somatosensory cortex. Journal of Neurophysiology 75:1589-1600 (1996)
Of course. They teach that as revealed dogma in kindergarten these days. So don't speak of a random process as if it were some sort of being as you did in the original passage. Nature does this, nature does that, bullshit that could all be replaced by "shit happens". But that doesn't sound nearly as plausible as "as the waters receded, nature made the fishies grow lungs" or "Biological Science tells us that nature worked this way in evolving intelligence". Intelligence happened. We have not a clue how. Concienceness happened. We have not a clue what it is, or how it came into being. Random process? Divine intervention? ET? There is basically no evidence much for any of them.
Is intelligence and self awareness inevitable? Common? Rare? Where is everyone?
Go lick a picture of Jerry Garcia. You are sounding like a Tim Leary wanna be.
I listened to the show. They didn't talk much about the blue brain stuff. More about intelligence and consciousness in general. They had Marvin Minsky and Paul Davies, among others as guests. It wasn't bad for a show aimed at the general public. The moderator tried to get people to talk about the "moral issues", but mostly they didn't.
So I still don't see what is so new about Blue Brain, especially compared to the existing parallel cortical simulation software such as NCS (NeuroCortical Simulator) that runs on Beowolf clusters and, if an offhand comment on their web site is to be believed, on Blue Gene.
formatting link
Note that Brainlab is a front end to simplify setting up models for NCS.
--
- Stephen Fuld
e-mail address disguised to prevent spam
One of your MIT friends did a PhD on self directed learning of speach. As a side effect, he had a data compressor that is better than gzip.
Paper is on xxx, sorry I can't remember the name. This was 96-97 or so.
--
Paul Repacholi 1 Crescent Rd.,
+61 (08) 9257-1001 Kalamunda.
West Australia 6076
comp.os.vms,- The Older, Grumpier Slashdot
Raw, Cooked or Well-done, it\'s all half baked.
EPIC, The Architecture of the future, always has been, always will be.
RM: "I don't think anyone really knows what consciousness is, never mind how to get something like a computer to be conscious."
GS: I think that there are a good many people who know what "consciousness" is. First, such people know that the word has a variety of meanings (i.e., its "use" is caused by different things at different times) and, second, they know that one use of "consciousness" involves our acquired ability to observe and talk about our own behavior. That is, essentially, what there is to say. When you understand that perception is behavior, and what we observe when we report our own perceptual behavior is that behavior, the problem of "qualia" disappears as well.
That makes sense, and simplifies matters. As long as each processor simulates more than one neuron, then, there probably will not be any connections from a processor to other than its nearest neighbors.
However, diagonally adjacent neighbors would have to be included as well, and even that is probably "distant" in the computer-science sense.
So now we know the ideal topology for a computer that simulates the brain: each processor has to have links to its *twenty-six* nearest neighbors in 3-D.
Oy. You guys have just re-invented the cellular automaton, or something like that. Not hardly the brain.
If you want a better idea of brain anatomy + operation, take a look at any book on the subject, and/or Jeff Hawkins new book On Intelligence, or any of the 6 or 8 books Gerald Edelman has written over the past
10-12 years [more accurately, he's re-written the same book that many times]. Look at the diagrams Edleman has of the massive#, as well as importance, of long-range fiber tracts. Check out the term "white matter".
I finally got a chance to listen from the streaming archive
formatting link
Worth the time, IMHO, just to hear what people are willing to say with a straight face. It seems like a fairly realistic presentation, except that there is really very minimal emphasis on the enormous gap between current capability (incredibly primitive) and the kinds of ethical questions they want to explore. Some reasonably incisive comments about self-organization and complexity.
The important ingredient is the corporate commitment and the marketing muscle of IBM. Whenever something interesting happens in AI, IBM wants to be there and wants to be known as a player.
I think people might all calm down and see this a little better if they realized that the name of the simulation is Blue BRAIN not Blue MIND. The scientists doing this do not pretend they are studying the mind (of course one cannot speak for the PR people and journalists); they are interested in simulating, right now, a very tiny structure in the brain consisting of about 10,000 neurons. (Not just 10,000 random neurons ala neural net. I don't know much brain anatomy, but apparently these 10,000 neuron elements make up distinctive elements of the brain that are replicated to form the larger brain.)
Obviously 10,000 neurons is not a full human brain. Obviously 10,000 neurons is also more brain than you get in certain animals --- clearly more than in a C. Elegans, and perhaps more than in a fairy fly.
If everything goes well, and if the calculations appear to provide useful and intriguing results, this will be scaled up to cover interacting versions of these 10,000 neuron elements.
The point, however, is not to simulate consciousness; it's to see if anything surprising and useful comes out of treating the brain as a very large neural net. If you believe Penrose there are limits to what will happen bcs an important aspect of the brain, some sort of quantum related stuff, is not being simulated. If you believe Timothy Leary types, there are limits to what will happen bcs brain chemistry is not (as far as I know) being simulated. Beyond brain chemistry there are plenty of other things that are not being simulated --- nothing about protein synthesis, nothing about cytoskeleton rearrangement. Simply stating these issues makes part of the point of the exercise a little more clear --- what CAN come out of such a simulation (and thus is dependent only on neural connections) and what, if anything, appears unable to come out of such a simulation?
And, of course, way before consciousness, there are plenty of more trivial brain functionalities we barely understand; things like vision and hearing and manipulating muscles in real-time to fly.
Yes, that a column, but it has more on the order of 50k-100k neurons. And the same thing holds: we can't even put it to a black box test, as we don't have any really useful input-output data for such a column. Be aware that such a model has on the order of 20-1000 degress of freedom _per_neuron_, so you are poking in a very highly dimensional space for a set of states that does "something useful".
An insect brain is on the order of a million neurons. That would at least make it possible to do useful tests. And C. elegans neural system certainly is also a very valid target, as not only are neurons but also all synapses are known.
Just as there is an enormous gap between our incredibly primitive current ideas about ethics and the problems they are purported to solve. For example, Brian Cantwell Smith (who was one of the participants) and I raised the 'thou shalt not kill' issue. If everyone made a daily backup copy, then murder would become a mere misdemeanor. What should be the punishment for causing someone to lose a half-day's work-considering that, with maintainable parts, that person already is potentially immortal.
Um, in fact, perhaps the punishment for 'killing' say, two days of backups should be more than that for killing the person itself.
Yes, as with Chess. Their publicity about Deep Blue said almost nothing about how they began by taking (or buying?) Deep Thought from Carnegie-Mellon.
============= Next month, IBM is set to activate the most ambitious simulation of a human brain yet conceived. It's a model they say is accurate down to the molecule. .... =============
"... molecule ..." !!
That certainly indicates why it might take a supercomputer node to model each neuron. Interesting that they think this level of detail is what it takes to do the job. Someone mentioned that, if Moore's law holds, then it will only be a few years before they can have a supercomputer node for each of the 10B-100B neurons, up from 10K with present system. Unless, of course, they decide by then that they need to push the models even further down ... say to the level of quantum effects on the atoms in the molecules. Oy.
Hi Jan. This is a good point. You also made a very good point earlier in the thread ....
============= I don't think you can easily distinguish primates's brains from different species just looking under the microscope...I believe that you'd have difficulty even with rat and human tissue. That's the whole point of mammalian neocortex: it's the quitessential information processing engine, as far as we can tell, and it's only the way it's wired to peripherals,
as well as its interactions with its peers (other parts of neocortex) that seem to determine its function. =============
The brain is a system of billions of elements which interact with each other, and it appears that it's really the interconnections between the many more-or-less "separate" processing areas [modules] which are the key to intelligence. The cortical columns are the best conjecture for local processing/computational units, but due to the neuroanatomical similarity across regions, it appears likely that the different regions engage in fundamentally the same types of operations internally.
The key point is that each area receives different input and output streams, and they have different interconnection pathways from one to the other. Edelman has a picture in his books which show the cat cortex has about 64 cortical areas with approx 1100 *major* interconnection pathways between them [this is the white matter]. The key to intelligence/etc probably lies in the complex nature of the signals flowing in these 1100 pathways, rather than in internal computations, per se, within the 64 areas. Edelman's ideas regards consciousness rests on the idea that the 1100 pathways essentially synchronize and integrate activity across the entire cortex at any point in time.
Also, regards the issue of stability, I started a thread on c.a.p a couple of years ago regards ... how can such a complex system ever be stable? Unfortunately, I don't recall too many useful responses to the question. The basic problem is that, even in simple "engineering" systems [say just 1 negative-feedback pathway], it is sometimes difficult to get good stability over many operating conditions. Now, take a system with billions of elements and feedback pathways like the brain, and the feedback is both positive and negative in variety. How can you EVER make such a system stable?
Personally, I believe that a much less detailed simulation must be work- able. Our brain has evolved to work in the face of so many varying boundary conditions, not to mention the physical and chemical harm we throw at it regularly, that its continued well-being (mental illnesses notwithstanding) cannot depend on such details of its physical structure. In my view, that's a large part of the interest in such work: What are the principles behind building such a complex system in such a way that it remains "stable" (in some fuzzy sense) in the face of fluctuations in its components? If we could better understand how that is done, we might be able to solve similar problems, e.g., how to build more stable societies or economies.
That makes the presuposition that a person is no more than the state of the organic materials that comprise their body. Many are unwilling to agree to that, in the total absence of evidence. Of course we all recall the "lime pit" thought experiment, in which travel is accomplished by transmitting the above hypothesized "backup data" and destroying the original. If one has enough bandwidth to record it routinely then one can transmit it and not worry about the bother of dealing with the actual carbon based matter. So, how many readers of this group would use such a means of transportation?
As I recall, they hired at least part of the CM team that did deep thought. Those folks then did deep blue as a follow on.
Interesting. Thanks for the comment. I can understand why stability "just happens" as an emergent property of natural selection, since any unstable mutations will no doubt die off quite rapidly, and those remaining will be characterized by being stable.
However, what would be the nature of "loose-coupling" in a system with
100B neurons and 100T synapses? How do we actually connect all of those cells together in a practical manner to attain this goal, rather than just saying that loose-coupling theoretically implies stability?
Take it from me. Running a four minute mile is not easy. And no one just steps onto a track and runs a four minute mile without significant preparation and training--even today.
And it doesn't. Interaction with an outside world brings forth intelligence.
Your comment trivializes the process of scientific and technlogical discovery in an embarrassing way. People might just get lucky at the horse track or in a singles bar, but not at the cutting edge of science and technology, where insight almost always precedes discovery. I say "almost" even though I can't offhand think of exceptions.
We're already seeing some important limitations in computers based on silicon. Those limitations are concealed by the numbers that are most often used to announce supercomputers in the press.
By one measure, only about 512 processors of genus Blue Brain work together effectively in problems requiring global communication, no matter how many processors you wire up to make a "supercomputer.".
The whole idea of using a one-dimensional measure to characterize computers as more or less powerful is profoundly misleading.
Blue Gene computers can do calculations that no single human or group of humans could ever do. On the other hand, Blue Gene couldn't pass the Turing test--not necessarily because it isn't "powerful" enough, but because no one knows how to program it to pass the Turing test.
ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here.
All logos and trade names are the property of their respective owners.