1

I have a number of records where I am trying to predict one field based on other fields. I set up a FANN neural net under Python, with ~10 inputs, 100 hidden nodes, and 2 outputs.

When I went to build the nets, with 1000+ records, they went down within a few thousand epochs to an error of about 60, but didn't go down any further; often the error reported was identical down to the last digit from the previous error.

When I went to test it out, I was expecting a weak correlation, but it predicted identical results for each piece of test data.

Now is this behavior because I'm not using FANN correctly (I hope), or because I'm working in a problem space where FANN is not particularly helpful?

ATdhvaankcse,

Christos Hayward
  • 5,777
  • 17
  • 58
  • 113
  • 5
    Sounds like you overfitted on the test data. Trained for too long/put the threshold too low. Also I think that 100 hidden nodes is very much considering your number of inputs and training set size. – Junuxx Jun 18 '12 at 14:30
  • Please describe how many possible answers exist, or if continuous the range of possible answers, and how you have mapped the two outputs to them. – wberry Jun 18 '12 at 15:57
  • Also, shouldn't there be an additional n in "ATdhvaankcse"? – Junuxx Jun 18 '12 at 16:22

0 Answers0