[eref]

> High representation efficiency

Capsules do require much fewer parameters, they generalize 10-20% better to new viewpoints, they are much more robust to adversarial examples and can better recognize overlapping objects; but, on the other hand, capsules currently require much more training data to achieve the same performance, even though in theory (if they would actually learn inverse graphics) they should require less data, and they add a lot of expensive additional structure (roughly 10X). I am rather pessimistic about whether the approach will lead us anywhere; it seems sub-optimal to model all possible child-parent configurations explicitly. That has a quadratic nature to it and my hunch is that can be done sub-linearily.