0
$\begingroup$

Some networks won't work without bias inputs. Is it necessary to acknowledge them in gradient descent and change them to minimize error? Or make just constant bias for every neuron and leave it that way?

I am making a huge net 256:64:32 for letter recognition and am hoping that it won't increase learning time by much.

$\endgroup$
0
$\begingroup$

The main function of the bias is to shift the activation function, which in turn helps in the learning process.

Have a look at this answer, I think it will fit your question as well : Role of Bias in Neural Networks

Hope this helps.

$\endgroup$

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.