Adding extra weights in NN


Have got some thoughts about artificially introduce some bias factor as extra weights in NN, in every layer of NN or the last one, is that feasible? how to do that?

What exactly do you mean by “bias factor”?

To add a ‘bias factor’ as extra weights I guess you can intercept the forward propagation in the linear transformation which is W * X + b.

Lets say you want to introduce a bias factor called W2.

Then you could probably go something like:

(W * W2) X + b

And then in the backprop you may want to consider this additional ‘constant’ in your derivatives.

Now, I really can’t predict what the result of this would be and/or how you would calculate that bias factor. Is that something you’ve probably elaborated already? can you share your thoughts?

Thank you for your reply. This is only a primary thought for CNN, sometime we want to boost some special class / groups of images. How can we know the exact information extracted in each layer of CNN? or can we use the result generated from CNN and use it in a second NN network?