Loading ...
Sorry, an error occurred while loading the content.
 

Re: Pruning When? and By how much?

Expand Messages
  • P. Dickens
    This contribution is probably usless but at least it is a contribution. No contribution, is useless, any contribution is useful if if the contributor believes
    Message 1 of 12 , Mar 2, 2007
      This contribution is probably usless but at least it is a
      contribution.

      No contribution, is useless, any contribution is useful if if the
      contributor believes it so, because even a negative can expose a
      positive and a Black Hole helps to gauge that something out there is
      causing a nothing. That's why young children are so good at saying
      "Why?", and by the time they reach sixteen, most of them have had it
      beaten out of them, which is about the biggest negative mankind faces,
      never mind Global Warming, and that one sits right under our noses.
      What the world needs are Einsteins that forever ask, Why? and get the
      answer.

      I regard your letter as one of those Why's?

      WELL DONE.


      Finally there is the ignorance factor. Those of us who might wish to
      offer a little insight or contribution but as we are only certain of
      our own ignorance might feel exposed if we open ourselves to
      criticism on an open forum.

      Of course that dose not apply to me. My ignorance is tempered by my
      arrogance. I am the fool who rushes in where the Angles (Angels - sic)
      fear to tread
      and I have the scars to prove it.

      I previously had a boss, who was extremely highly qualified
      internationally (though this was not generally known), and at any
      meeting, seminar or conference he had an inane ability to ask the most
      simple questions which to those working under him or around him would
      set looks of complete disbelief, a sort of "don't you know that?" or
      "how the hell can you ask that? (are you stupid or something)", yet it
      would almost always evoke answers from the recipient that were entirely
      unexpected by the rest of the audience, and especially those that had
      looked in disbelief when the question was asked, in the first place.
      In effect completely flooring everybody, that thought they could not ask
      such a stupid or simplistic question. Furthermore the answers were
      always telling or thought provoking. What I realised after some months
      of these experiences was that, what he was doing was asking the
      questions, just as a Child would do, and in many cases was doing it for
      the benefit of those around him. In truth he knew the answer already,
      which follows another managerial maxim, " Never ask a question, unless
      you know the answer ". It was a useful and educating experience from
      which a great many people benefited without realising it. A very useful
      way to side-step Pride and Arrogance or "Pride and Prejudice".


      It is people like me who wonder why cutting back the apple trees and
      roses is going to improve artificial intelligence, and would
      greatfully benefit, if possible, from a short explanation of what
      pruning is and why you would do it. This might not further your cause
      but would assist your silent audience.


      Your reference to Roses and Apples is very apt. Neural Networks (NN)
      are based upon the biological, in this case the Brain. They try to
      imitate methods that the biological world has developed through
      evolution over aeons (epochs). There are Neurones, Axons, Dendrites and
      Synapses in the brain. These are component parts of the functions of
      the Brain's activity, each perform a function that is part of an electro
      chemical chain, which I won't go into here. But suffice it to say that
      the Neuron is the key component that Neural Networks make use of and
      attempt to copy. When we are born the brain is a mass of neurons which
      are excited by stem cell growth and this continues up to the age between
      3 and 7. However, if the brain is not excited by learning and new
      knowledge along with the answers given to the incessant, but vitally
      important Why?, Why?, Why?, that a child will ask, just so long as they
      are given answers and not told to "Shut up", those Neurons will
      activate. If they don't they die off and it is a fact that a young
      child can have lost as many as 35 to 40% of it's Neurons that it
      originally had as a baby. This in effect scars a child for life and at
      the Pole end of this are adults in late life whose brain starts to die
      off, again through lack of use, (hence the Japanese idea of giving
      computer games to elderly people to reexcite those Neurons, and stop
      them dying off). You might say retirement is a death warrant (Brain
      Death). Playing Cards, Chess and Draughts will achieve the same
      objective as computer games. Teaching a baby two or more languages from
      birth is the best favour you can bestow on your child (e.g. in double or
      quad ethnic language families), because formative Language Neurons are
      in a different part of the brain to scholastically learned language
      Neurons.

      This whole process whether in a Baby cum Child or an Adult is " Pruning
      " . What the Brain doesn't use it will loose, as per the American
      proverb "Use it or Loose it.".
      Within a Neural Network. Your columns are your Nuerons each cell within
      that column representing a seperate Neuron. Each line of data
      represents a different Fact or Experience (as in life). We call them
      Facts. The Neural Network links each Neuron to every other Neuron.
      Yours does the same and the brain will compute all of the relevant
      experiences that it has, and the Dendrites and Axons, which form the
      interconnections between your Neurons will transmit the Electro chemical
      signal to each other Neuron going through an open connection called the
      Synapse, this open connection is bridged by a Morphine created by the
      brain for a millisecond, rather like a spark plug ( Hence, Drugs [e.g.
      Speed, Cocaine, Hash etc.] damage the Neurons by sealing these Synapses
      making them permanently linked and people then go GaGa). Within the
      Neural Network this duplicated part of the process occurs when
      calculations are done between each Neuron emulating what I have just
      described, as it occurs in the brain. As more and more of these
      calculations are done some of the links in the Neural Network between
      some Neurons become very small or even zero, now in Nature ' If it ain't
      used it is lost', i.e. it is "Pruned".

      Pruning therefore becomes a component part of emulating Brain activity.
      The question then is when to do it ? It is easy to say prune Zero
      connections, but in reality they are actually pretty rare even if the
      link is only 0.0000000001. Also the first set of calculations may not
      be the best time to Prune, because after a couple of Cycles (i.e.
      recalculation), that low value connection can suddenly increase. So in
      EasyNN pruning is set to take place after every 20 cycles (the
      default). The consequence of that pruning event is that low value
      connections i.e. Axions and Dendrites are removed, Neurons can thus
      become isolated and in effect be killed off, just as in the Brain. The
      question arises then, at what "Value" point to start Pruning i.e. at
      0.01 (the default value) or less or more ? I have also found that 20
      cycles is not always the best point to prune. Also when to start
      Pruning, do you start pruning at age 3, 4, 5, or 6 or in Neural network
      terms at 60% , 70%, 80% or 95% or more to my way of thinking at 67.36%
      or 87.45%, in other words these figure are purely arbitrary. The
      literature says "later" in the Neural Networks life i.e. when it nears
      Resolving or Converging, what is "Later", suggestions of 95, 97, and
      99%. I am not convinced by this. Nor am I convinced by the total Prune
      effect. i.e. once cut that's it. In nature if Neurons have been
      obliterated in early life it has been found that new ones can be
      encouraged to grow, not to the same extent, but they can and do regrow.
      My tests have found that Pruning can be instigated, which may then nudge
      a Network towards Resolving, when it has become stuck, but that
      reinstating the Neural connections after this nudge (by reinstating all
      connections again) can have a progressive effect. And then restart
      Pruning from scratch again, when it appears to be making little
      progress. The key element within all of this is that Pruning can be
      used to some good effect if a Neural Network stops making any progress.
      My original questions were aimed at finding some criteria and parameters
      that could be tried and tested.

      I know that there are Doctors, Engineers and Financial Analysts out
      there using this software and I know they will have had this problem,
      what I find breathtaking in the extreme, is their unwillingness to share
      their experiences here! Do they also tell their children to "Shut up "
      or "Stop keep asking stupid questions ", when their child keep asking,
      Why?


      And then there are the ignorant masses such as myself who read with
      interest the news letters. Have been through all of the demo
      exercises on NN. Have a vague Idea that neural networks should be
      real interesting, but cant think of a relevant application in my own
      existence / experience and are unable or unwilling to invest the time
      in setting up a NN problem just for the experience. Most of the
      problems I come across are solvable by specific solutions or are one
      offs that are better suited to the neural net between my ears.

      Solving problems is something our brain does all of the time, hence our
      ability to survive. However we are so neurointensive and psycho
      sexually driven, that our ability to focus on a single event (by ' event
      ' I mean problem, activity, repetitive action ) is extremely difficult,
      and we quickly become bored. Hence as children we are told to "Pay
      attention", "Could do better". These daydreams are a vital part of our
      neuro functions and not to be derided, it is what has got us to the Moon
      and back. The use of Neural Networks enables us to use a tool that
      would frankly be intolerable and would take a lifetime to solve, if done
      manually as a singular problem. Using other methods of problem
      resolution are there for the taking, NN is just one of them. Its
      advantage is that it can handle seemingly unconnected pieces of data and
      data in different formats (i.e. text, numerical, boolean). In our
      brain we can handle unusual connections for example between a Banana and
      having other food that is not yet ready to eat, because it is unripe.
      The Banana gives off a gas, which when placed with other unripe food
      will hasten its ripening. These represent a series of Biological and
      Chemical processes which in themselves are entirely unconnected. But
      our Brain can connect them and use them in a way that a Statistical test
      cannot, but similarly a NN can. That's its value.

      You may not have a work problem nor a domestic problem that warrants the
      use of a NN, but that doesn't mean there isn't something around what you
      do or have as an interest, that if enough of the right data was
      collected, could provide you with a direction, if not a solution, which
      having got there, could then well invoke those famous last words: " Now
      why didn't I think of that before ?"


      Thanks, Andrew for having the goodness of heart to respond.

      Also my thanks to Stef, whose point is only too well taken.

      P. Dickens





      Andrew wrote:

      > --- In easynn@yahoogroups.com, Stef <dataminer@...> wrote: At 09:33
      > 29/10/2007, you wrote:
      > > >Let us assume there are 1,000 Easy NN users out there, and that 3
      > % ( a
      > > >conservative and not unreasonable sample estimate) have used
      > Pruning,
      > > >that means that there are 30 of you out there, that have used
      > Pruning, if
      > > >there are 10,000 Easy NN users then that would make 300 that have
      > used
      > > >Pruning, who have used
      > > >Pruning, even if only once.
      > > >
      > > >Now let us say that only 50% of those figures (Users), are reading
      > this,
      > > >that still
      > > >leaves 15 or 150 users, who have used Pruning at some point in
      > time.
      > > >
      > > >Don't you think you could just send in a reply having now read
      > this ?
      >
      > >
      > > You now need to also factor in the general reluctance of discussion
      > > group members to "get involved". A great majority are just here to
      > > read, not write.
      > >
      > > This drastically reduces your projected figures, probably down as
      > low
      > > as one. Maybe that one is you?
      >
      > And then there are the ignorant masses such as myself who read with
      > interest the news letters. Have been through all of the demo
      > exercises on NN. Have a vague Idea that neural networks should be
      > real interesting, but cant think of a relevant application in my own
      > existence / experience and are unable or unwilling to invest the time
      > in setting up a NN problem just for the experience. Most of the
      > problems I come across are solvable by specific solutions or are one
      > offs that are better suited to the neural net between my ears.
      >
      > It is people like me who wonder why cutting back the apple trees and
      > roses is going to improve artificial intelligence, and would
      > greatfully benefit, if possible, from a short explanation of what
      > pruning is and why you would do it. This might not further your cause
      > but would assist your silent audience.
      >
      > Finally there is the ignorance factor. Those of us who might wish to
      > offer a little insight or contribution but as we are only certain of
      > our own ignorance might feel exposed if we open ourselves to
      > criticism on an open forum.
      >
      >
      > Of course that dose not apply to me. My ignorance is tempered by my
      > arrogance. I am the fool who rushes in where the Angles fear to tread
      > and I have the scars to prove it.
      >
      > This contribution is probably usless but at least it is a
      > contribution.
      > A
      >
      >


      [Non-text portions of this message have been removed]
    • P. Dickens
      Resent, Yahoo not picking up Italic of original text of original letter. ... contribution. No contribution, is useless, any contribution is useful if if the
      Message 2 of 12 , Oct 3, 2007
        Resent, Yahoo not picking up Italic of original text of original
        letter.

        >>>This contribution is probably usless but at least it is a
        contribution.

        No contribution, is useless, any contribution is useful if if the
        contributor believes it so, because even a negative can expose a
        positive and a Black Hole helps to gauge that something out there is
        causing a nothing. That's why young children are so good at saying
        "Why?", and by the time they reach sixteen, most of them have had it
        beaten out of them, which is about the biggest negative mankind faces,
        never mind Global Warming, and that one sits right under our noses.
        What the world needs are Einsteins that forever ask, Why? and get the
        answer.

        I regard your letter as one of those Why's?

        WELL DONE.
        ---

        >>>Finally there is the ignorance factor. Those of us who might wish to
        offer a little insight or contribution but as we are only certain of
        our own ignorance might feel exposed if we open ourselves to
        criticism on an open forum.

        >>>Of course that dose not apply to me. My ignorance is tempered by my
        arrogance. I am the fool who rushes in where the Angles (Angels - sic)
        fear to tread and I have the scars to prove it.

        I previously had a boss, who was extremely highly qualified
        internationally (though this was not generally known), and at any
        meeting, seminar or conference he had an inane ability to ask the most
        simple questions which to those working under him or around him would
        set looks of complete disbelief, a sort of "don't you know that?" or
        "how the hell can you ask that? (are you stupid or something)", yet it
        would almost always evoke answers from the recipient that were entirely
        unexpected by the rest of the audience, and especially those that had
        looked in disbelief when the question was asked, in the first place.
        In effect completely flooring everybody, that thought they could not ask

        such a stupid or simplistic question. Furthermore the answers were
        always telling or thought provoking. What I realised after some months
        of these experiences was that, what he was doing was asking the
        questions, just as a Child would do, and in many cases was doing it for
        the benefit of those around him. In truth he knew the answer already,
        which follows another managerial maxim, " Never ask a question, unless
        you know the answer ". It was a useful and educating experience from
        which a great many people benefited without realising it. A very useful
        way to side-step Pride and Arrogance or "Pride and Prejudice".
        ---


        >>>It is people like me who wonder why cutting back the apple trees and
        roses is going to improve artificial intelligence, and would
        greatfully benefit, if possible, from a short explanation of what
        pruning is and why you would do it. This might not further your cause
        but would assist your silent audience.

        Your reference to Roses and Apples is very apt. Neural Networks (NN)
        are based upon the biological, in this case the Brain. They try to
        imitate methods that the biological world has developed through
        evolution over aeons (epochs). There are Neurones, Axons, Dendrites and
        Synapses in the brain. These are component parts of the functions of
        the Brain's activity, each perform a function that is part of an electro

        chemical chain, which I won't go into here. But suffice it to say that
        the Neuron is the key component that Neural Networks make use of and
        attempt to copy. When we are born the brain is a mass of neurons which
        are excited by stem cell growth and this continues up to the age between

        3 and 7. However, if the brain is not excited by learning and new
        knowledge along with the answers given to the incessant, but vitally
        important Why?, Why?, Why?, that a child will ask, just so long as they
        are given answers and not told to "Shut up", those Neurons will
        activate. If they don't they die off and it is a fact that a young
        child can have lost as many as 35 to 40% of it's Neurons that it
        originally had as a baby. This in effect scars a child for life and at
        the Pole end of this are adults in late life whose brain starts to die
        off, again through lack of use, (hence the Japanese idea of giving
        computer games to elderly people to reexcite those Neurons, and stop
        them dying off). You might say retirement is a death warrant (Brain
        Death). Playing Cards, Chess and Draughts will achieve the same
        objective as computer games. Teaching a baby two or more languages from
        birth is the best favour you can bestow on your child (e.g. in double or

        quad ethnic language families), because formative Language Neurons are
        in a different part of the brain to scholastically learned language
        Neurons.

        This whole process whether in a Baby cum Child or an Adult is " Pruning
        " . What the Brain doesn't use it will loose, as per the American
        proverb "Use it or Loose it.".
        Within a Neural Network. Your columns are your Nuerons each cell within
        that column representing a seperate Neuron. Each line of data
        represents a different Fact or Experience (as in life). We call them
        Facts. The Neural Network links each Neuron to every other Neuron.
        Yours does the same and the brain will compute all of the relevant
        experiences that it has, and the Dendrites and Axons, which form the
        interconnections between your Neurons will transmit the Electro chemical

        signal to each other Neuron going through an open connection called the
        Synapse, this open connection is bridged by a Morphine created by the
        brain for a millisecond, rather like a spark plug ( Hence, Drugs [e.g.
        Speed, Cocaine, Hash etc.] damage the Neurons by sealing these Synapses
        making them permanently linked and people then go GaGa). Within the
        Neural Network this duplicated part of the process occurs when
        calculations are done between each Neuron emulating what I have just
        described, as it occurs in the brain. As more and more of these
        calculations are done some of the links in the Neural Network between
        some Neurons become very small or even zero, now in Nature ' If it ain't

        used it is lost', i.e. it is "Pruned".

        Pruning therefore becomes a component part of emulating Brain activity.
        The question then is when to do it ? It is easy to say prune Zero
        connections, but in reality they are actually pretty rare even if the
        link is only 0.0000000001. Also the first set of calculations may not
        be the best time to Prune, because after a couple of Cycles (i.e.
        recalculation), that low value connection can suddenly increase. So in
        EasyNN pruning is set to take place after every 20 cycles (the
        default). The consequence of that pruning event is that low value
        connections i.e. Axions and Dendrites are removed, Neurons can thus
        become isolated and in effect be killed off, just as in the Brain. The
        question arises then, at what "Value" point to start Pruning i.e. at
        0.01 (the default value) or less or more ? I have also found that 20
        cycles is not always the best point to prune. Also when to start
        Pruning, do you start pruning at age 3, 4, 5, or 6 or in Neural network
        terms at 60% , 70%, 80% or 95% or more to my way of thinking at 67.36%
        or 87.45%, in other words these figure are purely arbitrary. The
        literature says "later" in the Neural Networks life i.e. when it nears
        Resolving or Converging, what is "Later", suggestions of 95, 97, and
        99%. I am not convinced by this. Nor am I convinced by the total Prune
        effect. i.e. once cut that's it. In nature if Neurons have been
        obliterated in early life it has been found that new ones can be
        encouraged to grow, not to the same extent, but they can and do regrow.
        My tests have found that Pruning can be instigated, which may then nudge

        a Network towards Resolving, when it has become stuck, but that
        reinstating the Neural connections after this nudge (by reinstating all
        connections again) can have a progressive effect. And then restart
        Pruning from scratch again, when it appears to be making little
        progress. The key element within all of this is that Pruning can be
        used to some good effect if a Neural Network stops making any progress.
        My original questions were aimed at finding some criteria and parameters

        that could be tried and tested.

        I know that there are Doctors, Engineers and Financial Analysts out
        there using this software and I know they will have had this problem,
        what I find breathtaking in the extreme, is their unwillingness to share

        their experiences here! Do they also tell their children to "Shut up "
        or "Stop keep asking stupid questions ", when their child keep asking,
        Why?
        ---

        >>>>And then there are the ignorant masses such as myself who read with
        interest the news letters. Have been through all of the demo
        exercises on NN. Have a vague Idea that neural networks should be
        real interesting, but cant think of a relevant application in my own
        existence / experience and are unable or unwilling to invest the time
        in setting up a NN problem just for the experience. Most of the
        problems I come across are solvable by specific solutions or are one
        offs that are better suited to the neural net between my ears.

        Solving problems is something our brain does all of the time, hence our
        ability to survive. However we are so neurointensive and psycho
        sexually driven, that our ability to focus on a single event (by ' event

        ' I mean problem, activity, repetitive action ) is extremely difficult,
        and we quickly become bored. Hence as children we are told to "Pay
        attention", "Could do better". These daydreams are a vital part of our
        neuro functions and not to be derided, it is what has got us to the Moon

        and back. The use of Neural Networks enables us to use a tool that
        would frankly be intolerable and would take a lifetime to solve, if done

        manually as a singular problem. Using other methods of problem
        resolution are there for the taking, NN is just one of them. Its
        advantage is that it can handle seemingly unconnected pieces of data and

        data in different formats (i.e. text, numerical, boolean). In our
        brain we can handle unusual connections for example between a Banana and

        having other food that is not yet ready to eat, because it is unripe.
        The Banana gives off a gas, which when placed with other unripe food
        will hasten its ripening. These represent a series of Biological and
        Chemical processes which in themselves are entirely unconnected. But
        our Brain can connect them and use them in a way that a Statistical test

        cannot, but similarly a NN can. That's its value.

        You may not have a work problem nor a domestic problem that warrants the

        use of a NN, but that doesn't mean there isn't something around what you

        do or have as an interest, that if enough of the right data was
        collected, could provide you with a direction, if not a solution, which
        having got there, could then well invoke those famous last words: " Now
        why didn't I think of that before ?"
        ----

        Thanks, Andrew for having the goodness of heart to respond.

        Also my thanks to Stef, whose point is only too well taken.

        P. Dickens

        ----------------------------



        Andrew wrote:

        > --- In easynn@yahoogroups.com, Stef <dataminer@...> wrote: At 09:33
        > 29/10/2007, you wrote:
        > > >Let us assume there are 1,000 Easy NN users out there, and that 3
        > % ( a
        > > >conservative and not unreasonable sample estimate) have used
        > Pruning,
        > > >that means that there are 30 of you out there, that have used
        > Pruning, if
        > > >there are 10,000 Easy NN users then that would make 300 that have
        > used
        > > >Pruning, who have used
        > > >Pruning, even if only once.
        > > >
        > > >Now let us say that only 50% of those figures (Users), are reading
        > this,
        > > >that still
        > > >leaves 15 or 150 users, who have used Pruning at some point in
        > time.
        > > >
        > > >Don't you think you could just send in a reply having now read
        > this ?
        >
        > >
        > > You now need to also factor in the general reluctance of discussion
        > > group members to "get involved". A great majority are just here to
        > > read, not write.
        > >
        > > This drastically reduces your projected figures, probably down as
        > low
        > > as one. Maybe that one is you?
        >
        > And then there are the ignorant masses such as myself who read with
        > interest the news letters. Have been through all of the demo
        > exercises on NN. Have a vague Idea that neural networks should be
        > real interesting, but cant think of a relevant application in my own
        > existence / experience and are unable or unwilling to invest the time
        > in setting up a NN problem just for the experience. Most of the
        > problems I come across are solvable by specific solutions or are one
        > offs that are better suited to the neural net between my ears.
        >
        > It is people like me who wonder why cutting back the apple trees and
        > roses is going to improve artificial intelligence, and would
        > greatfully benefit, if possible, from a short explanation of what
        > pruning is and why you would do it. This might not further your cause
        > but would assist your silent audience.
        >
        > Finally there is the ignorance factor. Those of us who might wish to
        > offer a little insight or contribution but as we are only certain of
        > our own ignorance might feel exposed if we open ourselves to
        > criticism on an open forum.
        >
        >
        > Of course that dose not apply to me. My ignorance is tempered by my
        > arrogance. I am the fool who rushes in where the Angles fear to tread
        > and I have the scars to prove it.
        >
        > This contribution is probably usless but at least it is a
        > contribution.
        > A
        >
        >


        [Non-text portions of this message have been removed]
      • Andrew
        ... % ( a ... Pruning, ... Pruning, if ... used ... this, ... time. ... this ? ... low ... And then there are the ignorant masses such as myself who read with
        Message 3 of 12 , Nov 2, 2007
          --- In easynn@yahoogroups.com, Stef <dataminer@...> wrote: At 09:33
          29/10/2007, you wrote:
          > >Let us assume there are 1,000 Easy NN users out there, and that 3
          % ( a
          > >conservative and not unreasonable sample estimate) have used
          Pruning,
          > >that means that there are 30 of you out there, that have used
          Pruning, if
          > >there are 10,000 Easy NN users then that would make 300 that have
          used
          > >Pruning, who have used
          > >Pruning, even if only once.
          > >
          > >Now let us say that only 50% of those figures (Users), are reading
          this,
          > >that still
          > >leaves 15 or 150 users, who have used Pruning at some point in
          time.
          > >
          > >Don't you think you could just send in a reply having now read
          this ?

          >
          > You now need to also factor in the general reluctance of discussion
          > group members to "get involved". A great majority are just here to
          > read, not write.
          >
          > This drastically reduces your projected figures, probably down as
          low
          > as one. Maybe that one is you?

          And then there are the ignorant masses such as myself who read with
          interest the news letters. Have been through all of the demo
          exercises on NN. Have a vague Idea that neural networks should be
          real interesting, but cant think of a relevant application in my own
          existence / experience and are unable or unwilling to invest the time
          in setting up a NN problem just for the experience. Most of the
          problems I come across are solvable by specific solutions or are one
          offs that are better suited to the neural net between my ears.


          It is people like me who wonder why cutting back the apple trees and
          roses is going to improve artificial intelligence, and would
          greatfully benefit, if possible, from a short explanation of what
          pruning is and why you would do it. This might not further your cause
          but would assist your silent audience.


          Finally there is the ignorance factor. Those of us who might wish to
          offer a little insight or contribution but as we are only certain of
          our own ignorance might feel exposed if we open ourselves to
          criticism on an open forum.


          Of course that dose not apply to me. My ignorance is tempered by my
          arrogance. I am the fool who rushes in where the Angles fear to tread
          and I have the scars to prove it.

          This contribution is probably usless but at least it is a
          contribution.
          A
        • Andrew
          P. Dickens wrote: No contribution, is useless, any contribution is useful if if the ... Thank you for your kind words of wisdom upon pruning,
          Message 4 of 12 , Nov 5, 2007
            "P. Dickens" <pdickens@...> wrote:
            No contribution, is useless, any contribution is useful if if the
            > contributor believes it so.

            Thank you for your kind words of wisdom upon pruning, and your
            spelling correction.

            Concerning the postulation of a problem siutable for NN I was
            interested to read your discussions on such a subject.
            It seems to me that the main two areas where NN's are being applied
            are in an attempt to predict the random. Eg predicting the output of
            a horse race or footbal match or other gambling event, or the future
            of the stock exchange.

            Predicting the future of the stock exchange is probably relativly
            easy if you could monitor the several million factors upon which it
            is dependant. While the outcome of a horse race probably has much
            fewer dependancies, getting the infomation upon them is probably
            equaly impossible without "inside information" on every horse. Wheras
            for football matches the number of unpredictable random events
            influencing the outcome is probably too high to get a reliable
            predictor. The people using NN for this are making the assumpion that
            the events they are tryin to model are predictable or significantly
            influenced (ststistialy influenced)by the data they are using as
            inputs. They could equaly well be using astronomical charts. And I
            bring to mind the swis proffessor who based on 25 years reserch found
            a rank correlation coefficient of over 0.99 between the number of
            nesting storks in Switserland and the number of births. He was of
            course smart enough to know that these statisics can only disprove a
            link not prove one.

            I had thoulght of a problem such as rating photos and then getting NN
            to scan through the rest of my photos and predict how I would rate
            them? I could then adapt the hetrustic to allow other people to rate
            the photos and then again search for the photos that most appeal to
            them.


            Is it posible to get more than one output from NN? Could I get NN to
            look at photos and tell me weather there was an animal in the photo?
            what type of animal there was? wether there were people in the photo?
            how many? Was Jane in the photo? Was Jack in the photo? ...etc.

            Once again please excuse my ignorance. And kindly forgive an spelling
            mistakes I may have made as I am chalanged in that department.
          • P. Dickens
            ... Yes, in theory you can specify as many outputs as you wish. This however becomes increasingly less desirable. There is in fact no black and white answer to
            Message 5 of 12 , Nov 5, 2007
              >>>Is it posible to get more than one output from NN? <<<<<<

              Yes, in theory you can specify as many outputs as you wish.
              This however becomes increasingly less desirable.
              There is in fact no black and white answer to this one.
              Often the perceived wisdom is wrong.
              However one cannot be dogmatic about this because the origins of that
              perceived wisdom may have proveded substantive evidence to support it in
              the first place, but all to often following it can lead you uyp the
              garden path. I was once told some years ago that because I was inputing
              a range of data pertainig to two variable that I should only request of
              the NN two output variables I wanted. This after months of work was
              just not working and having asked early on about just getting one output
              variable instead of two, that that would not work, and that was the
              perceived wisdom . Guess what, when I decided to tell the person
              concerned that they were probably wronmg I proceded to do it my way and
              guess what chasing one output at a time but each in turn. eventually
              gave me the results I was looking for,, and they were wrong. However
              they had a good many more years experience and in all cases that they
              had been involved with and that ran to thousands, they had been proved
              right, so I cannot say that that person was wrong they were just wrong
              when it came to the data we were dealing with at that time, so this is
              nothing to gloat about, it is just part and parcel of the way NN's work,
              " the Black Box ".
              --------


              >>>Could I get NN to look at photos and tell me weather there was an
              animal in the photo?
              what type of animal there was? wether there were people in the photo?
              how many? Was Jane in the photo? Was Jack in the photo? ...etc.<<<<<<<

              The answer is yes, in each case.
              NN dedicated software is used on Airport and Port recognition ssytems.
              It is also used on data and script recognition systems mainly sold to
              Local Authorites, Government Ministries and Large companies with
              specialised needs. .
              For speed, dedicated NN hardware (processors) already preprogramed are
              used.
              What you are suggesting would be quite an undertaking.
              You would need to start with one face or animal and get that to work.
              Steve has done a lot of work on EasyNN to enhance its image handling
              cpabilities in recent months and years.
              ---------


              >>>>Concerning the postulation of a problem siutable for NN I was
              interested to read your discussions on such a subject.
              It seems to me that the main two areas where NN's are being applied
              are in an attempt to predict the random. Eg predicting the output of
              a horse race or footbal match or other gambling event, or the future
              of the stock exchange. <<<<<<<


              Certainly NN are used to attempt to deal with the Random but results may
              well be unsatisfactory.
              The whole point about Rondomness is just that, it is Random and not
              predictable, because being random it has no linearity nor regressive
              behaviour. It does not even fit into a Poisson distribution. Random is
              more closely associated with Chaos. However evn in chaos htere may well
              be patterns which are discerenable but the scale upon which this
              distinguishing pattern is so enormas and may well involve more than
              three dimensions that it is impossible for use to assimilate anything
              usefull. hence trying to use NN to predict the Lottery is waste of time
              except that you have a 1 in 13,983,816 chance of getting the right set
              of numbers, so the NN has got one chance of getting it right !.

              Predicting in the areas you mention are not based upon Randomness, but
              upon variables which may or may not have a bearing upon the output.
              Certainly NN can and is used to that effect. NN are used in Medicine
              for example to identify problem X Ray images that are then passed to a
              Surgeon. Also as a diagnostic tool. In property it can be used for
              Valuations. In another case it is used for acurately identifying a
              specific type of Wool and used by a Wool trader (Importer, Wholesaler)
              to reduce the risk of expensive mistakes. The list is prettywell
              limitless.
              ---------


              >>>>.......and I bring to mind the swis proffessor who based on 25 years
              reserch found
              a rank correlation coefficient of over 0.99 between the number of
              nesting storks in Switserland and the number of births.<<<<<<<

              There is a simple but profound axiom in Statistical Research:-
              "Correlation does not mean Causation."
              ---------

              >>>>I had thoulght of a problem such as rating photos and then getting
              NN
              to scan through the rest of my photos and predict how I would rate
              them? I could then adapt the hetrustic to allow other people to rate
              the photos and then again search for the photos that most appeal to
              them. <<<<<<<<<<

              The rating of images would provide you with a level of complexity that
              could be mindblowing, if not 'Interesting'
              Since you wiould have to create a rating sytem based on multiple
              variables, which would then be each placwed on scalar template.
              Your Variables would also have to be used across multiple catagories,
              and many would be intrinsically emotionally based (such as grading
              faces), whilst others would be value judgments (such as height, or warm
              and sunny, Amateurish or Profesional looking image) Yet others could be
              factual (such as Black and White or Sepia or Colour) and so on.
              You would need to do several test samples, (as they do in Market
              research, test the questionnaires first, e.g. If you have to explain a
              question, it is NO good.), before going full steam ahead.
              An interesting Project, which would be well worth writing up when done.

              P. Dickens

              ----------------
              Andrew wrote:

              > "P. Dickens" <pdickens@...> wrote:
              > No contribution, is useless, any contribution is useful if if the
              > > contributor believes it so.
              >
              > Thank you for your kind words of wisdom upon pruning, and your
              > spelling correction.
              >
              > Concerning the postulation of a problem siutable for NN I was
              > interested to read your discussions on such a subject.
              > It seems to me that the main two areas where NN's are being applied
              > are in an attempt to predict the random. Eg predicting the output of
              > a horse race or footbal match or other gambling event, or the future
              > of the stock exchange.
              >
              > Predicting the future of the stock exchange is probably relativly
              > easy if you could monitor the several million factors upon which it
              > is dependant. While the outcome of a horse race probably has much
              > fewer dependancies, getting the infomation upon them is probably
              > equaly impossible without "inside information" on every horse. Wheras
              > for football matches the number of unpredictable random events
              > influencing the outcome is probably too high to get a reliable
              > predictor. The people using NN for this are making the assumpion that
              > the events they are tryin to model are predictable or significantly
              > influenced (ststistialy influenced)by the data they are using as
              > inputs. They could equaly well be using astronomical charts. And I
              > bring to mind the swis proffessor who based on 25 years reserch found
              > a rank correlation coefficient of over 0.99 between the number of
              > nesting storks in Switserland and the number of births. He was of
              > course smart enough to know that these statisics can only disprove a
              > link not prove one.
              >
              > I had thoulght of a problem such as rating photos and then getting NN
              > to scan through the rest of my photos and predict how I would rate
              > them? I could then adapt the hetrustic to allow other people to rate
              > the photos and then again search for the photos that most appeal to
              > them.
              >
              > Is it posible to get more than one output from NN? Could I get NN to
              > look at photos and tell me weather there was an animal in the photo?
              > what type of animal there was? wether there were people in the photo?
              > how many? Was Jane in the photo? Was Jack in the photo? ...etc.
              >
              > Once again please excuse my ignorance. And kindly forgive an spelling
              > mistakes I may have made as I am chalanged in that department.
              >
              >


              [Non-text portions of this message have been removed]
            Your message has been successfully submitted and would be delivered to recipients shortly.