11

im having a response like below from fann

    Epochs            1. Current error: 0.2500066161. Bit fail 4.
    Epochs           58. Current error: 0.0000930788. Bit fail 0.

what does Bit fail mean here?

ibaguio
  • 2,298
  • 3
  • 20
  • 32

3 Answers3

6

The bit fail limit is the maximum difference between the expected and actual output neuron value that is allowed.The default bit fail limit is 0.35. If the difference between the expected and actual output neuron value is more that the bit fail limit, this counts as 1 bit fail. In the sample output you gave, at 58 epochs all the output neurons gave actual outputs close enough to the expected outputs and hence the bit fail was 0 and training stopped. In other words all the training examples gave outputs that were close enough to the expected outputs. During the first epoch, 4 of the training samples gave outputs resulting in bit fails.

Junk Mvg
  • 157
  • 2
  • 7
  • If the output should be either 0 or 1, should the bit fail limit be set to zero? I'm also confused how the bit fail limit works with "desired error". – krasnaya Jan 24 '15 at 16:02
4

from documentation of FANN

The number of fail bits; means the number of output neurons which differ more than the bit fail limit http://leenissen.dk/fann/html/files/fann_train-h.html#fann_get_bit_fail

Birol Kuyumcu
  • 1,195
  • 7
  • 17
  • 2
    what does it mean when it says the number of output neurons differ? i don't think i understand the concept behind it. can someone explain or point to where i can read about it? – webnat0 Jul 12 '13 at 23:54
2

Yea I have found this confusing as well and thought that it may have been a bug in 'ruby-fann'.

The FANN manual states that it is the number of output neurons failing but doesn't say that it is the total sum of the number of output neurons for the provided sample set. Therefore the worst case 'Bit fail' is ALL of the output neurons failing (beyond the specified bit fail limit) for ALL of the samples.

Doug
  • 21
  • 2