What is the difference between considering and ignoring bias?

Hi :)

I'm studying for explaining DNN and have some questions for considering bias and batchnormalization in source code.

What is the exact difference between keeping bias and ignoring bias?

Is it the process of subtracting the effect of bias from the input relevance?

or is it process for only considering the activation of neuron before and after applying bias?

And for explaining models with batchnormalization (resnet, densenet), is it fully implemented? or still processing?

It seems that the heatmaps provide better explanations when algorithm ignores the batchnormalization reverse process and just propagates input relevance to next layer.

Thank you for reading.

From Nam