C4 W1 MobileNet, can number of computations be reduced further?

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

Hello!
I am on this slide right now (I hope you can see the screenshot - it is on 10:44, when we compute number of operations for pointwise multiplication).

image
image1344×698 93.7 KB

It seems like when we calculate the first layer of the result stack (on the right) we are using 2 vectors: 1x1x3 from the input and 1x1x3 from the filter.
When we calculate the second layer, we again use the same 2 vectors.

So as far as I can see all the layer of the output will be the same as they are calculated using the same vectors.
The question is: can we skip the multiplications we do not need and just copy the values from the first layer?

2

Maybe I’m just missing your point, but note that the filter values are different for each output layer. E.g. in that instance there are 5 different 1 x 1 x 3 filters, so you need to do everything 5 times.

3

Oh, that is true @paulinpaloalto - their initial values are random, and every filter has unique behavior. This explains the idea. Thank you for clarifying this for me!