Initialization of Weights in Logistic and Linear Regression

Dear Professors,

I went through the FAQs, and still, I was not able to understand why we can use Zero initializations in Logistic Regression and we can’t use Linear Regression. Please if you could explain it once more?

Thank you for your help!!

Did you read this thread? That was linked from the FAQ Thread. If the explanations there don’t make sense to you, then please ask a more specific question about what was said on that thread.

The really short version of the answer is: try it and see what happens. When you use Zero Initialization in Logistic Regression it clearly works. When you try it with the network in Week 3 Planar Data exercise, it clearly does not work to use zero initialization. Try it and see for yourself. Gradient Descent can’t learn anything and the weights just stay zero. The thread I referenced explains in mathematical terms why that happens. It’s all math.