6
$\begingroup$

I am building a neural network. Each perceptron in the network uses a sigmoid activation function.

Must I normalize my inputs (which currently range form 0 to 1200)? I ask this because the sigmoid function approaches 1 as the input to it approaches infinity. Hence, if my input is not between 0 and 1, I'm afraid that the sigmoid will always return 1.

$\endgroup$
6
$\begingroup$

The inputs should be scaled to the so-called "active range" of the activation function, or, in other words, the area of the function curve where the derivative of the function is clearly non-zero. This is done for backpropagation to work properly, since it uses activation function derivatives, and ~ 0 derivatives imply extremely small (insignificant) changes to NN weights (no learning). For sigmoid, the active range lies somewhere between -sqrt(3) and sqrt(3). You may scale inputs to that range.

Also, yes: sigmoid will always output values in (0;1), because that's sigmoid's range. You will need to scale NN outputs to the necessary ranges.

$\endgroup$

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.