Variable Resolution Network
A network that can be put in front of any temporal network architecture. A time series might have areas of low resolution -not much happening for many succesive time steps- followed by a short, high resolution section. The former benefits from longer kernels while the latter needs shorter kernels. This pre-network tries to learn how to combine kernels of different lengths by incorporating a spatial step after an initial temporal step. As far as I'm aware, this isn't done yet. Typically, firstly, a spatial layer might be used, sometimes called a bottleneck layer, to reduce the complexity of the input data, in the case of a multivariate data source.