As far as I know choosing an activation function for the input layer is relatively straightforward: I use Sigmoid if the input data domain is (0,1) and TANH if it is (-1,1).

But what activation functions to set for hidden and output layers? Is there any conventional logic for making thish choice reasonably? How do I know/set the domain of a neuron layer output?

  • Input: it is not common to use activation functions in the input layer. Just rescale your data to the [-1;1] interval;
  • Hidden: the $tanh$ activation used to be the most popular. Now, this role has been taken by the $relu$ (rectified linear unit) activation, which produce sparse activations and better preserve the gradients;
  • Output: here is where you need to chose the activation function according to your problem type. For classification use the softmax activation (the multivariate version of the logistic sigmoid). For regression problems you may use linear outputs (identity activation function).
| cite | improve this answer | |

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.