I'm new to machine learning and am trying to write a simple neural network that uses back-propagation. Now, so far I've successfully implemented my neural network to learn a boolean function. So for example, I've tested it on the AND, OR, XOR, IFF and some random 3-input boolean function. It's managed to learn all of them so far. The next step I wanted to take was to do numeric prediction. Now I have an input data set, but I have no idea on what the differences between my current implementation and the numeric prediction implementation would be. I just barely have my head wrapped around the whole idea of back-propagation so I'm a little confused on what I would need to change.

My current implementation is written in Java.

Just to recap real quick, I basically have a neural net that can learn boolean functions. I want to modify this to be able to make numeric predictions. Is this an easy task, and how would I go about this? From my understanding, I don't see why my current implementation wouldn't work out.

Please let me know if you need any more information. Thank you!

  • $\begingroup$ Maybe you didn't normalize your input and target data? Probably you use sigmoid activation layer in backpropagation network and there output will be always between 0 and 1 (or -1 and 1 for hyperbolic tangent layer) $\endgroup$ – itdxer May 22 '15 at 17:24
  • $\begingroup$ @Globmont have you finished course on Stanford about machine learning? $\endgroup$ – 404pio May 22 '15 at 17:29
  • $\begingroup$ Sorry for the late responses. I'm based in Australia :) @itdxer Yeah, I didn't normalize it. How would I go about normalizing it then getting it back to it's original scale? $\endgroup$ – Globmont May 23 '15 at 4:30
  • $\begingroup$ @frankov Nope, I haven't done the Stanford course. I'm actually in a course right now in uni. Is the Stanford course very good? $\endgroup$ – Globmont May 23 '15 at 4:30
  • $\begingroup$ Check this answer stats.stackexchange.com/questions/70801/… $\endgroup$ – itdxer May 23 '15 at 8:47

Replace the activation function (logistic sigmoid?) with the identity function. Your cost function should also be swapped to MSE (in case you are using cross entropy or neg log likelihood). The rest should be basically the same.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.