Quantifying and Learning Static vs. Dynamic Information in Deep Spatiotemporal Networks