/RouteNet

Primary LanguagePython

RouteNet

Goal: Feedforward neural network that has hard gates such that features are passed from one layer to the next in a content-aware fashion, reducing the overall number of computations needed for inference. There is some relation to CapsuleNets--however, those networks use iterations to perform dynamic routing from a capsule in layer i to a single target capsule in layer i+1, whereas the desired network here has no recurrence/iteration, and routing that can send features in one layer to more than one destination in subsequent layers.

Notes

Similar ideas are finally starting to show up in the literature. E.g.,