1. Limited time only! Sign up for a free 30min personal tutor trial with Chegg Tutors
    Dismiss Notice
Dismiss Notice
Join Physics Forums Today!
The friendliest, high quality science and math community on the planet! Everyone who loves science is here!

The Brain

  1. Oct 3, 2003 #1
    Is it possible for us to make a functional model of the human brain?
    I'm not interested in the phylosophical implications, but more in the resources, techonolgies we can use (that's why I posted here.... )
     
  2. jcsd
  3. Oct 3, 2003 #2

    Nereid

    User Avatar
    Staff Emeritus
    Science Advisor
    Gold Member

    Why not start small?

    How about a honeybee? Or a simple worm?

    If those could be modelled, wouldn't the human brain just be a matter of scaling up?
     
  4. Oct 3, 2003 #3
    hi Nereid :smile:

    I don't think insects have a central nervous system (I could be wrong though...).
    What animal has the simplest brain; is there a biologist here that can answer that?
    just found this link here: http://www.artificialbrains.com/ for anybody interested
     
  5. Oct 3, 2003 #4

    Njorl

    User Avatar
    Science Advisor

    Not quite.

    Mammalian brains are different from reptiles and birds, which are different from fish which are different from invertabrates.

    I think a human brain can be modelled. It is a physical entity after all. However, not all of its functionality is understood, and what is understood is tremendously complex.

    Njorl
     
  6. Oct 3, 2003 #5

    Nereid

    User Avatar
    Staff Emeritus
    Science Advisor
    Gold Member

    What are the key differences?

    What is the simplest mammal, from the brain perspective?
     
  7. Oct 3, 2003 #6

    Njorl

    User Avatar
    Science Advisor

    I'm no expert, and would screw up the details I'm sure. Try this link:

    http://primatesociety.com/Into/survival/timeline/textEvol.html

    I'm not qualified to judge the info on the site. Hopefully, one of our biology experts will critique the ideas presented at this site, and maybe come up with a better one if they find it lacking.

    Njorl
     
  8. Oct 6, 2003 #7
    What if you wrote a physics engine in a computer language and programmed in every atom of a human brain? :)
     
  9. Oct 7, 2003 #8
    well I was hoping of something that could be done in my lifetime...
     
  10. Oct 7, 2003 #9
    Unless you're extremely old, it could! All we need is a very fast computer with a ton of memory, and we can program a program that sets up every atom, after detecting it in a machine with a dead human brain (or some such thing) in it.

    Something like that, anyway. It's not as impossible as it sounds...
     
  11. Oct 7, 2003 #10
    (pay no attention to the man behind the curtain!)


    Guybrush,

    It is most definitely not possible. There is, at this time, a worldwide effort underway to simply map the human brain. In other words they still haven't sorted out what functions different parts of the brain perform.

    If you take vision, for example, they know this is primarily processed in the occipital lobes, but important aspects of vision, such as the perception of up and down motion, are processed in the parietal lobes. Other aspects are processed in the temporo-parietal area. What is one phenomenon to us: vision, is actually the perception of many different aspects of vision fused into the one we're conscious of.

    No one even thinks to consider we need a system to process the perception of up and down motion till it gets damaged in someone.
    We don't know how many things the brain does that we haven't even realized needed doing.

    The functions performed by the brain are, therefore, probably still not all discovered, and it is taking imput from neurologists all over the world, to simply create this comprehensive map.

    That being the case it is clear they don't know remotely enough to create a functioning model of the circuitry.

    If you'll settle for a good plastic brain model go here:

    Brain Mart's Brain Models
    Address:http://www.brain-mart.com/brain_models.html [Broken]

    Zoobyshoe
     
    Last edited by a moderator: May 1, 2017
  12. Oct 7, 2003 #11

    FZ+

    User Avatar

    Ah yes, I raised this in the chat, didn't I?

    I think the whole issue is that of acceptable approximation. It is not required that we model every atom etc, but rather just understand what each individual cell or group of cells does and translate this into an entire system. If there are parts of the brain we don't know about... then modelling them would be outside the scope of our model's criterias.

    In using GR as a model of the universe, we do not claim that GR covers all things we do not even know. Rather, we judge the acceptability of the model on what we already know...
     
  13. Oct 8, 2003 #12
    IMO the great challenge is to make a system that can learn. No system made by man can do that yet. It may not be obvious but computers work on algorithms. Every algorithm in use today was made from scratch by men. There isn't a algorithm that was made by a computer. I think that if we want to make a model of the brain sooner or later we must design a system that can learn...

    sikz:
    the human genome project lasted 13 years and al they did was mainly:
    - identify all the approximately 30,000 genes in human DNA,
    - determine the sequences of the 3 billion chemical base pairs that make up human DNA.
    Tenths of institutes have joined resources to make this happen. (http://www.genome.gov/ )
    And you want ot map every atom in the brain and believe it can be done in a lifetime? You're too optimistical :smile:

    zooby:
    that's a start as good as any other...
     
  14. Oct 8, 2003 #13

    Njorl

    User Avatar
    Science Advisor

    Modelling the brain in isolation is not really a model of the brain as it works in humans. There is significant bio-feedback from chemicals produced in other parts of the body. For example, eyes detect tiger, tiger is referenced in brain, brain tells adrenal glands to produce adrenaline, adrenaline affects neurotransmission, brain works differently.

    Njorl
     
  15. Oct 8, 2003 #14
    Guybrush wants a functioning model, a very tall order, and one which means a physical apparatus that does something, as opposed to a conceptual model, such as "Bohr's model of the atom".

    If we were to make a functioning apparatus based on what we do know about the brain, leaving out the parts we don't know about, it really wouldn't be a model of a human brain at all. With luck it might amount to a functioning model of a very damaged human brain.

    Njorl's post points out the that the functioning of the brain as we know it extends out and into the body. I hadn't thought of that. Guy's model would actually require
    a functioning model of the body to be built as well, to support it.
     
  16. Oct 9, 2003 #15
    u can only create a probabilistic model using AI algorithms like genetic algo and neural networks, which themselves are derived from biological principles on which our brain and nervous system works. so, the fuzziness in our final model will definitely limit the efficiency, and we will end up with a brain that does not, or rather, can not ,THINK ....
     
  17. Oct 9, 2003 #16
    model atoms of brain?

    Modelling the atoms may be too ambitious, but modelling the individual functions could be also. The power contained in the brain comes from the fact that all of these functions interact synergistically, so that the whole is more than the sum of the parts, and we may never be able to determine all of the unseen interactions that take place when several functions combine to produce a thought. Rather, let's consider the middle road. Just model the connections between neurons, and the conditions that cause each synapse to fire (kind of like the game of life). I picture an AI containing groups of synapses of various characteristics - in this group, two inputs (on average) cause a firing while in that group it may be that 5 inputs are required. As stated, chemical processes can change the characteristics of the neurons, but such chemical processes would be fairly easy to model as a possible output of certain groups having high activity at the same time, and this causing a global change throughout the AI, with various affects to each group. In summary, I believe the synergistic characteristics of the brain come mainly from the circular feedback of various neuron groups each having DIFFERENT firing rules on average. Also important would be if one group, for example, comes between a pair of groups in the normal communication scheme, i.e. the corpus colosum connects the hemispheres, but modifies the info going between them.
    Just some random thoughts on the subject.
    Aaron
     
    Last edited: Oct 9, 2003
  18. Oct 10, 2003 #17
  19. Oct 13, 2003 #18
    feedback

    I'm picturing the brain as being multiple neural networks, having feedback between many pairs of such networks. In both the brain and in all other complex biological, chemical, and physical systems that exhibit synergistic effects and emergent phenomena, feedback is the key. Without feedback, things go in a linear fashion. With feedback, surprises can result! Picture a corporation running with information and direction going only from the top down - disaster! The management will be out of touch with what is going on below them, and even in parallel branches of the corporation. Those who implement policies will have no recourse when conditions change and the directions they were given no longer fit the situation they now find themselves in. Parallel branches may act against the goals of each other - a schizophrenic corporation. Now add upward and sideways feedback and voila, we now have an intelligently run corporation. It's the same way with the brain. In simplistic terms, at the very least we have feedback between memories, reasoning, input, planning, output, emotions (intuitions, hunches), feelings, imaginings, etc. How many different facets of thought are there - I'd like people to supply their own answers to this here, if they think it would help the thread. Now, how many of these facets are stand-alones, don't give or receive feedback from other facets during normal thinking?
    Aaron
     
    Last edited: Oct 13, 2003
  20. Oct 13, 2003 #19
    Aaron,

    From what I understand of the brain you are on the right track.
    All the different parts of the brain are constantly in communication with other parts, back and forth.

    I would only suggest that for the "multiple neural networks" of the first sentence of your last post you substitute "multitudinous neural networks." Earlier today I was looking through a book on the brain and became aware of yet more things the brain does that I hadn't ever considered needing doing.

    You have also to take into account the all important, constant stream of new information coming in through the sences. The more I consider the brain/body/environment interaction the less I think it is possible to construct even the crudest of functioning models.
     
  21. Oct 21, 2003 #20

    turin

    User Avatar
    Homework Helper

    What do you mean by "learn?" In the simplest, zeroth order sense, is not any response to stimulus learned?
     
  22. Nov 28, 2003 #21

    Yes what exactly is the defenition of learn? It is not just simply retaining information. ok looked it up. To gain knowledge, comprehension, or mastery of through experience or study.
    To fix in the mind or memory; memorize: learned the speech in a few hours.

    To acquire experience of or an ability or a skill in: learn tolerance; learned how to whistle.
    To become aware: learned that it was best not to argue.
    To become informed of; find out. See Synonyms at discover.
    Nonstandard. To cause to acquire knowledge; teach.
    Obsolete. To give information to.

    So some of this stuff a computer can do, like memory and processing. but to cause a computer to become aware or to have a drive to discover its environment on its own, hmmmmmmmm. I am contemplating Is this differant than processing, hmmmmmmmm lets see, I think in order to make a computer learn it would need to become aware of itself. hmmmmmmm I am computer therefor I am. We designed the algorithms for the computer to use so what do we use is the question? if not complex algorithms than what? We want to learn, why? If a computer did not know it exsisted than why would it want to learn? It wouldn't. I don't care how much memory or power you give it, it will still be a machine that can just simply crunch information and spit it back out at you. If the computer senses its exsistance though watch out. Even at a low computational level it will learn. After that sight, feel, eye movement becomes more than a sensor, it becomes its connection to its environment, It will want to move a sight sensor and will start to figure out how or if it can, in order to see something that seems interesting. Now you've got something!
     
  23. Nov 30, 2003 #22

    Nereid

    User Avatar
    Staff Emeritus
    Science Advisor
    Gold Member

    If you know how it's happening, it's not learning?

    sheldon,

    There are plenty of examples from the insect world where learning - in the sense you describe above - doesn't happen, yet I'm sure you'd agree that the critters do each have a brain.

    In the world of complex computer systems, a good case could be made that some systems do 'learn', if you simply judge their behaviour in the same way as some behaviour in simple animals is called 'learning' (by the relevant biologists).

    Perhaps you don't count what the autopilot of a modern fighter plane, or the management system of a telephone network, can do as being able to 'learn' because you know (in principle) how the code was built?

    Sure, a wasp can't learn how to whistle or learn not to argue and neither can an autopilot. However, is there a fundamental difference between a wasp and, say, a chimp? Or is it just a matter of time before a meta-autopilot could behave like a chimp?
     
  24. Nov 30, 2003 #23
    Engineering answer, imo, not in few hundred years.

    There is still some uncertainty whether we could at all. Reason is complexity. There are limits to our human comprehension. Infact they are even measurable. For ex, it has been suggested to "just write a program", but its obvious difficulty of writing such program isn't even imagined. In past years some interesting revelations have came up. For eg, in 70ies they built a computer with 65536 cpus. It was disaster, not because it was slow (it was fastest supercomputer at a time), but because programmers of a time were simply unable to program it to utilise its potential. Occured that humans simply are unable to think parallel enough. Thats an important bummer. Technology was sitting there, and we couldn't make use of it.

    Today, its said that roughly, human being is able to cooperatively comprehend around 1 million lines of code. Beyond that, number of errors introduced outnumbers fixes made. For every 1 fix you add >1 error. Another important bummer. It seems to set the limits as to how complex a program we can create. MS windows is around 1M lines of code. Not very intelligent. Fortunately, the limit is not hard one, we evolve on that front.

    Then, there is interesting question whether a machine can at all comprehend its own design and functioning? Answer isn't trivial at all, because to comprehend its own functioning, it should be able to comprehend more than what it is designed for, in addition to that its designed for. We can dismantle brain into functional units and perhaps even understand how they work, but to put together coherent functional whole, we'd need to really make a monumental effort. We are far from even trying. We'll need to comprehend it as a whole. Would we have enough brain capacity for that?

    As to techonolgies, quantum computing seems to come close enough for the job. When it matures, and we learn to program it, we'll be able to build equivalent to brain. Then, we'll have to solve our ability to program something as complex as that. And I believe that will be the main obstacle. I like to think its possible, but it would take leap in structured cooperative programming approaches. And thats unfortunately very slow process, could take generations of geniuses to get us ready.
    It seems more probable that we'll create learning and evolutioning computers that would be better suited for such difficult tasks, and that they would then succeed in creating thinking machine. Thats beginning of a new species, homo roboticus
     
  25. Dec 4, 2003 #24

    Nereid

    User Avatar
    Staff Emeritus
    Science Advisor
    Gold Member

    Do you have a source with more data on this bug/program size conundrum?

    IIRC, there were >1m lines-of-code programs running quite well long before Windows came along. Also, to what extent have the 'must-not-have-bugs' apps run into a size problem of the kind you describe? I'm thinking of things like the latest Airbus and Boeing autopilot software, nuclear reactor control systems, medical lasers, and so on.
     
  26. Dec 5, 2003 #25
    No, I don't have a source handy. It was few years back when I saw it.

    Before windows, LOC were measured in tens of thousands.
    There is always a difficulty to define exactly what is meant by lines of code and program. NT4.0 was said to have over 9M LOC, w2000 over 40M. But thats not a single program. There is some essential part of it that makes up the "core" of it, thats much smaller. Its no point in including Notepad as a part of the core, as its pretty irrelevant to it.

    There is no need to look at whole system as one program. But there are interactions of a system that require good coordination of any changes in it, because parts depend on each other. Its those dependancies that grow into huge mess. By program is usually meant such conglomerate of tight interdependancies. By line of code is usually meant minimal conceptual unit of computer action that changes state of it. Not comments, not empty lines, not syntactic sugar.

    1Million lines of code is impossible for single person to even read through, let alone understand, and forget about remembering beginning while reaching the end. And there are no two persons that think the same. Many people working on same code are bound to misunderstandings, hurting each others work while "fixing" bugs of other, etc. More complex programs have more and deeper dependancies on its own internal state, so there is a trend - the more complex the program becomes, the more difficult it is to change anything in it without braking loads in other areas, or even find where the error is.

    There are alot of ways to manage a teamwork, and main leap is expected exactly there. OO programming is one example that made things possible unthinkable in the past. New ideas are coming that would extend the raw LOC limit, but there remains something essential to it. Its not that we should know exact value of the limit, but just that there is one, and its not very huge nor complex.

    As to autopilot, I don't think its very complex thing. Its difficult to invent and design, but the program code itself isn't very huge.
    I recall, that NASA had rules that accounted that comprehension matter. IIRC, deep space probes and other space computers had under 60K lines of code, had to be developed by single person, had to be developed concurrently by 5-10 persons independantly and unrelatedly, then exchanged code and had to understand and find out all errors other programmers made. Repeated over few times. Then best design picked. There is noone out there to press ctrl-al-del, y'know..

    I believe modern autopilots of planes also have 2-3 systems built by different coders, all computing same task, and then comparing with each other in realtime - trying to detect computing errors, too much is at stake. But given that airbus computers runs on windows, I don't know.. must be that it doesn't matter.
     
Share this great discussion with others via Reddit, Google+, Twitter, or Facebook