question from paper - Githubissues

facebookresearch / SparseConvNet

Submanifold sparse convolutional networks

https://github.com/facebookresearch/SparseConvNet

Other

2.06k stars 334 forks source link

question from paper #245

Closed hfutdzx closed 9 months ago

hfutdzx commented 9 months ago

hello，thank you for your contribution，but i have a question when i read your paper.

“hidden layers in the network may not receive all the information they require to classify the input data: in particular, two neighboring connected components are treated completely independently”

i can’t understand this sentence which are from your paper.

why do hidden layers treat neighboring connected components completely independently？

btgraham commented 9 months ago

'Submanifold convolutions' pass messages only amongst immediate neighbors, and only if they are active. Downscaling and then upscaling space with a UNet allows distant parts of the input to get signals from each other; once you downscale enough everything becomes connected.

hfutdzx commented 9 months ago

Thank you, I have understanded it. Components do not refer to pixel or point