Question about initialization Unit 4.2(Part 1-3)

[jihobak]

Unit. 4.2 Multilayer Neural Networks (Part 1-3)

Let's say activation func is relu and If we initialize weights and bias to zero, then all of hidden nodes have zero value after one forward propagation.

here's a question, Can I say all parameters(weights and bias) don't change even after backprop? because backprob is the change of multiplication and there's zero layer in the middle.

[rasbt]

Yeah, that's a fair point. Since all values are 0, even the bias units, there won't be an update since all gradient are 0