Hacker News new | comments | ask | show | jobs | submit login
A technical report on convolution arithmetic in the context of deep learning (github.com)
34 points by ghosthamlet 72 days ago | hide | past | web | favorite | 6 comments

How 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).

Extremely. If padding=valid, you are possibly throwing information away. If padding=same, you are possibly introducing invalid data that will throw things off.

Yes, 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.


Looks like all the outputs have smaller dimensions than the inputs. How would you go the other way?

See 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.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact