http://theorangeduck.com/media/uploads/other_stuff/phasefunction.pdf

In this network structure, the weights are computed via a cyclic function which uses the phase as an input.

Dynamically changing the network weights as a function of the phase instead of keeping them static as in standard neural networks significantly increases the expressiveness of the regression while retaining the compact structure. This allows it to learn from a large, high dimensional dataset where environmental geometry and human motion data are coupled.