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

easyNN sometimes doesn't learn

Expand Messages
  • jluhrsen
    Hi, I m using an older version of EasyNN (14.0g). Sometimes, when I try to start training, the learning is stagnant. There is no change in the error values no
    Message 1 of 4 , Oct 23, 2012
    • 0 Attachment
      Hi,

      I'm using an older version of EasyNN (14.0g).

      Sometimes, when I try to start training, the learning is stagnant. There is no change in the error values no matter how many cycles I
      allow.

      If I build a new network from scratch, it will usually start learning,
      even with the same settings. Sometimes toggling the trimming on/off/on will initiate it to a learning state. Sometimes, if I re-associate or "forget learning" in the GUI it will get it to learn.

      I can't figure out exactly how to get to this state, or exactly how to get out of this state.

      Any thoughts on this?

      I wanted to use the same data grid and build many nets with different numbers of nodes. After training one, I always seem to end up in a non-learning state for further nets.

      Thanks,
      James
    • P. Dickens
      I wanted to use the same data grid and build many nets with different numbers of nodes. After training one, I always seem to end up in a non-learning state for
      Message 2 of 4 , Oct 24, 2012
      • 0 Attachment
        I wanted to use the same data grid and build many nets with different
        numbers of nodes. After training one, I always seem to end up in a
        non-learning state for further nets.

        Are you merely substituting the data, that you already have in the grid,
        and that has resolved or converged, with new data?
        Or are you building a new net from scratch with the new data, but using
        the same grid structure as the previous net?

        OR

        Are you resetting the number of nodes in the parameters, when you build
        the net or more likely substituting the old data for the new, but at the
        same time failing to change the number of nodes (either increasing or
        decreasing them? If this is the case then you need to build a new net,
        setting the parameters accurately.

        P. Dickens


        From:
        "jluhrsen" <jamo@...>


        * Save Addresses
        <http://email10.orange.co.uk/webmail/en_GB/read.html?FOLDER=SF_INBOX&IDMSG=61665&check=&SORTBY=1#>

        To:
        easynn@yahoogroups.com
        Date: Oct 24 2012, 10:00 AM
        Subject:
        easyNN sometimes doesn't learn

        Show full header
        <http://email10.orange.co.uk/webmail/en_GB/read.html?FOLDER=SF_INBOX&IDMSG=61665&check=&SORTBY=1#>

        Hi,

        I'm using an older version of EasyNN (14.0g).

        Sometimes, when I try to start training, the learning is stagnant. There
        is no change in the error values no matter how many cycles I
        allow.

        If I build a new network from scratch, it will usually start learning,
        even with the same settings. Sometimes toggling the trimming on/off/on
        will initiate it to a learning state. Sometimes, if I re-associate or
        "forget learning" in the GUI it will get it to learn.

        I can't figure out exactly how to get to this state, or exactly how to
        get out of this state.

        Any thoughts on this?

        I wanted to use the same data grid and build many nets with different
        numbers of nodes. After training one, I always seem to end up in a
        non-learning state for further nets.

        Thanks,
        James



        [Non-text portions of this message have been removed]
      • SkywarnPB6027
        Doea anybody know how to expand the min or max value or force the Easynn to extend scale. I have new data that I need to add and test but the scale is too
        Message 3 of 4 , Oct 24, 2012
        • 0 Attachment
          Doea anybody know how to expand the min or max value
          or force the Easynn to extend scale. I have new data
          that I need to add and test but the scale is too high on
          the min scale .

          Does anyone know.? I know there is a feature to expand
          scale but that does not work well.. I need to physically
          put a lower number in there for the min values especiallty
          ones that are the predictors.

          Paul
          On 10/23/2012 2:18:00 PM, jluhrsen (jamo@...) wrote:
          > Hi,
          >
          > I'm using an older version of EasyNN (14.0g).
          >
          > Sometimes, when I try to start training, the learning is stagnant. There is no change in the error values no matter how many cycles I
          > allow.
          >
          > If I build a new network from scratch, it will usually start learning,
          > even with the same settings. Sometimes toggling the trimming on/off/on will initiate it to a learning state. Sometimes, if I re-associate or "forget learning" in the GUI it will get it to learn.
          >
          > I can't
          > figure out exactly how to get to this state, or exactly how to get out of
          > this state.


          [Non-text portions of this message have been removed]
        • jluhrsen
          Hi, to be more specific, let me give a recent example: I built and trained a 154 node network on 2300 examples. I saved the .tqv file and did another save
          Message 4 of 4 , Oct 25, 2012
          • 0 Attachment
            Hi,

            to be more specific, let me give a recent example:

            I built and trained a 154 node network on 2300 examples. I saved
            the .tqv file and did another "save as" to create my next net. Now,
            I changed the number of nodes to 175. After doing that, I clicked
            yes to "reset all learning". I kept the same control settings and
            let it re-optimize the rates which I think resets all learning again.
            Now, when I start the training all error values are flatlined
            and no learning happens. I stopped the learning and clicked to
            "forget all learning" and tried again. Still, no learning. Finally,
            I closed the program, reopened EasyNN and re-opened my new net. This time I set my min layer1 nodes to 175 and max to 176. I let it "pick"
            the number of nodes and it picked 176. learning was again reset and
            this time the net learned.

            I can't pinpoint when it will learn and when it wont. Maybe I'm just
            no understanding some concept?

            Thanks,
            James



            --- In easynn@yahoogroups.com, "P. Dickens" <pdickens@...> wrote:
            >
            > I wanted to use the same data grid and build many nets with different
            > numbers of nodes. After training one, I always seem to end up in a
            > non-learning state for further nets.
            >
            > Are you merely substituting the data, that you already have in the grid,
            > and that has resolved or converged, with new data?
            > Or are you building a new net from scratch with the new data, but using
            > the same grid structure as the previous net?
            >
            > OR
            >
            > Are you resetting the number of nodes in the parameters, when you build
            > the net or more likely substituting the old data for the new, but at the
            > same time failing to change the number of nodes (either increasing or
            > decreasing them? If this is the case then you need to build a new net,
            > setting the parameters accurately.
            >
            > P. Dickens
            >
            >
            > From:
            > "jluhrsen" <jamo@...>
            >
            >
            > * Save Addresses
            > <http://email10.orange.co.uk/webmail/en_GB/read.html?FOLDER=SF_INBOX&IDMSG=61665&check=&SORTBY=1#>
            >
            > To:
            > easynn@yahoogroups.com
            > Date: Oct 24 2012, 10:00 AM
            > Subject:
            > easyNN sometimes doesn't learn
            >
            > Show full header
            > <http://email10.orange.co.uk/webmail/en_GB/read.html?FOLDER=SF_INBOX&IDMSG=61665&check=&SORTBY=1#>
            >
            > Hi,
            >
            > I'm using an older version of EasyNN (14.0g).
            >
            > Sometimes, when I try to start training, the learning is stagnant. There
            > is no change in the error values no matter how many cycles I
            > allow.
            >
            > If I build a new network from scratch, it will usually start learning,
            > even with the same settings. Sometimes toggling the trimming on/off/on
            > will initiate it to a learning state. Sometimes, if I re-associate or
            > "forget learning" in the GUI it will get it to learn.
            >
            > I can't figure out exactly how to get to this state, or exactly how to
            > get out of this state.
            >
            > Any thoughts on this?
            >
            > I wanted to use the same data grid and build many nets with different
            > numbers of nodes. After training one, I always seem to end up in a
            > non-learning state for further nets.
            >
            > Thanks,
            > James
            >
            >
            >
            > [Non-text portions of this message have been removed]
            >
          Your message has been successfully submitted and would be delivered to recipients shortly.