Talk:Capsule neural network

I've finished with this piece now. It is ready for submission and/or publishing. Feedback encouraged. Lfstevens (talk) 02:02, 29 December 2017 (UTC)[reply]

In my version I had carefully split out intuition and motivation. As it is now the article seems like the whole idea emerged from nowhere. Those that has follwed the field for some time knows that this idea has a history long before the latest papers.

Now the article is about visual representation of objects, which is kind of weird. First of all, spatial relationship between parts of an object is not the object itself. This is about information represented in a vector space, and how transformations on the information can reconnect otherwise disparate points in that space. That can be used for reconnecting visual parts of an object, but that is only one of several use cases.

In my version I wrote about how it compares to cortical minicolumns, now it is about a special case where capsule networks are used for interpreting visual scenes. My version was about how capsule nets can be interpreted in the general case, now it is about a special solution for a special problem.

Nearly everything about the inspiration from cortical minicolumns are gone, which is very bad, as it removes the foundation for why capsule networks have some serious flaws. The flaws is how the dynamic routing is done, actually the error is the routing. Inspecting the history it is clear that you simply dismissed that as speculation. Well, to put it bluntly, capsule nets are not how the brain does it, but it is a small step in the right direction.

No, I don't think this draft should be posted as it is now. Jeblad (talk) 00:13, 30 December 2017 (UTC)[reply]