What is Veeky Forums's opinion on capsule networks proposed by Geoffrey Hinton?
Paper:
arxiv.org
Github code:
github.com
What is Veeky Forums's opinion on capsule networks proposed by Geoffrey Hinton?
Paper:
arxiv.org
Github code:
github.com
Other urls found in this thread:
people.idsia.ch
en.wikipedia.org
openreview.net
twitter.com
Just like everything else
>falling for the simplification meme
>implying it's necessarily tractable to provide guarantees for a model's accuracy
>implying regression validation isn't good enough for most practical applications
Randall sure is a retard
>What if the answers are wrong?
What does he mean by that? Is he talking about the known answers in the training sets? If you can't trust your training set data then you shouldn't be doing supervised learning in the first place. And if he instead means the answers generated by the program, you deal with them being wrong by having an error function to measure distance from the training set answers and the whole point is having it update to move in the direction of decreasing error.
It seems like a weird ensembling trick
It reminds me of the architecture in 'Multi-column Deep Neural Networks for Image Classification': people.idsia.ch
>It seems like a weird ensembling trick
it's sort of the opposite, from what i can tell.
What's the opposite of ensembling?
Its basically combining Hough transforms and neural nets. Its basically a coincidence detector.
Nose here, mouth here, eye here, other eye there = high probability of face.
The main difference is that the spatial relationships of those subcomponents are taken into account. In a normal convolutional net, simply the presence of a feature is what is used, whereas with a capsule the "pose" of a feature is corroborated with the "pose" of other features to do inference.
Hinton is 69 years old and still producing original research. Truly original, as in no one else was even thinking about this type of shit. Dude is a legend.