CNBC Enhancement Notes

• Enhancements incorporated in lens-2.64.cnbc1

  CNBC101

  • Replaced standard writeExample function with version allowing control over frequency of writing example information.
  • Added command setWriteFrequency which provides various options for customizing the activation output file format.
    The usage format is

     setWriteFrequency <frequency_value> [ -notarget ] [ -label ]  

    where <frequency_value> can be one of the following - ONCE_PER_TICK(default),ONCE_PER_INTERVAL or ONCE_PER_EXAMPLE which allows users to control how often output and target values are recorded during training or testing of network.Users can chooose to not have the target history written for each example by setting the -notarget flag. The -label flag when specified generates outputs labelled by group names and example names for better readability of activation file.

  CNBC102

  • Added the -u (use_average_error) flag in the train command to customize the format of the error report. The use of this flag enables the calculation and printing of average and maximum error values for every reporting interval. Alternatively reportAvgError flag can be set to get the same result using the setObject command ,

     setObj reportAvgError 1 

    will generate average and maximum errors for updates between reporting intervals in the network training report.

  CNBC103

  • Enhanced functionality of parallel training ,by dividing examples across clients according to load.(Thanks to Dave Plaut)
    Added a network object parallelSplit.This is used for parallel training when examples have to be assigned to clients in sets.e.g

    setObj parallelSplit 2 

    will ensure that examples are passed on to clients in multipes of 2.

  CNBC104

  • Added the -i flag in the noiseType command for inverting binary external input values for groups having noise function EXT_INPUT_NOISE.The command usage is -
    

      noiseType <group-list> [ -range <range> | -multiply | -flat | -invert <flip1> <flip0>] 

    where -invert option causes the external input values to be flipped ,for groups having noise type EXT_INPUT_NOISE. The flip factors <flip1> and <flip0> determine the probability by which external bits set to 1 and 0 should be inverted.

  CNBC105

  • Fixed a bug in the calculations for instant error derivatives of intergrating input and output units for continuous networks. (Thanks to Dave Plaut)

If anyone would like to share information about some generic customizations that they have made to their lens code, send a note to Sindy after inserting your documentation in this template.