Cost function added to backpropagation

Course Q&A Deep Learning Specialization Sequence Models
1

In first programming assigment from week 1, there is RNN with many-to-many relation, where Tx = Ty. I fully understand backprop that is provided in notebook, but I have no idea how Loss/Cost Function at each time step t should affect parameters - Waa, Wax, Wya, ba, ba, by. At the end of timestep, we calculate da_prev as shown on a figure below:

image
image913×533 41.6 KB

but should not da_prev be also affected by cost function of y^(t-1) ? If so, how to calculate “true da_prev”?

2

The loss function doesn’t directly influence the gradients.

It’s the partial derivative of the loss function that gives the gradients that are used in backpropagation.

3

Yes, you are right. So how the partial derivative of the loss function affects gradients?

4

Ok. The answer was in the picture. Total da and timestep t equals sum of da from cost function at t and da from future (t+1) timestep. Issue can be closed