Greetings!!
I am sorry I am unable to understand how overfeat algorithm works. In particular, (around 6:30 of the video), Andrew says “But it turns out a lot of this computation done by these four convnets is highly duplicative. So what the convolutional implementation of sliding windows does is it allows these four pauses in the convnet to share a lot of computation.” How can they share the same computation? every cell of the filter is either right-shifted or down-shifted by stride. How can the computations be same? What am I missing? Is there a resource for understanding the intuition? kindly let me know. thank you!
Which week of the course, and which lecture?
Course-4 Week-3 lecture titled “Convolutional implementation of sliding windows”
The computation that is the same is just the standard application of the convolutional filter as you describe. It steps through the image and does the same thing at each position, subject to the stride. The answer isn’t the same at each position because the inputs aren’t the same, but the coefficients of the filter are the same. So why do you need sliding windows? The “sliding” part is just what convolutions do.
ok, thank you. as I understand there is no reduction in computation per se. we still need to do (Nc x f x f) element-wise multiplications and (Nc x f) additions for each filter that is applied.
I understand without sliding, only one part of the image will be convolved with.