1

I have trained a neural network and am getting the accuracies in the end, but how do I go about getting the confidence in each prediction? The neural network is being used for classification, so I want to know how confident the neural network is that a certain input yields a certain output. So how do I get how confident the neural network is that a certain data point should be classified as 1 or 0? My neural network currently outputs values between 0 and 1, but they are not always rounded. So for example if one output is .6, then that rounds up to 1 but how do I obtain a confidence value from that?

In order to see which points the nn was unconfident on, I set a threshold such that I removed the data for which the neural network predicted between .4 and .6, since .5 is the middle and it could go either way to 1 or 0. If the nn predicted .5 then it is really unsure as to whether or not the data should be classified as 1 or 0 so this threshold removed those unconfident predictions

whuber
  • 281,159
  • 54
  • 637
  • 1,101
a13a22
  • 163
  • 12
  • I'm not sure the appended bit was meant as part of the question. It sounds like an answer. – Firebug Jul 25 '17 at 15:34
  • I posted it as the answer but I'm not sure why it was changed – a13a22 Jul 25 '17 at 15:36
  • Your intuition is correct that the NN is the least confident in the predictions that are close to 0.5 are the "least confident" predictions. But your decision rule to round up from 0.6 to 1 or down from 0.4 to 0.0 doesn't seem especially well-grounded. See the discussion here. https://stats.stackexchange.com/questions/127042/why-isnt-logistic-regression-called-logistic-classification/127044#127044 – Sycorax Jul 25 '17 at 16:06

0 Answers0