3 ms·
A technical report on convolution arithmetic in the context of deep learning
- amelius 8y agoHow important is padding in actual applications? I'm wondering if it's a detail one can just forget about without any consequences (except of course for matching dimensions in subsequent computations).
- therobot24 8y agovery
- tsomctl 8y agoExtremely. If padding=valid, you are possibly throwing information away. If padding=same, you are possibly introducing invalid data that will throw things off.
- amelius 8y agoYes, but it's information at the edge of the image. Show someone a photo of a cat, but clip off the edges, and they'll still recognize the cat.
- amelius 8y agoLooks like all the outputs have smaller dimensions than the inputs. How would you go the other way?
- joshvm 8y agoSee the dilated convolution further down the page. You spread the result of the convolution over a wider area and then the gaps are filled in as you scan over the rest of the image you're convolving.